Add additive *_with_rng entropy-injection variants to the Olm API

[shakhvit]

Summary

Adds additive *_with_rng variants to every key-generation entry point in the olm
module, letting callers supply their own impl CryptoRng instead of the thread-local
rand::rng(). Fully backwards-compatible: no existing method changes behaviour.

Motivation

vodozemac currently sources all key-generation entropy from rand::rng() (the
thread-local, OS-seeded generator), with no way to inject a specific RNG. Exposing an
RNG parameter is a standard, widely-useful capability that x25519-dalek and
ed25519-dalek already provide at the primitive level
(StaticSecret::random_from_rng, SigningKey::generate). Surfacing it at the
Account/Session level enables:

• Deterministic tests / KAT vectors — reproduce exact keys and ciphertext from a
  fixed seed, so regressions in key-derivation plumbing are caught byte-for-byte.
• Reproducible builds and record/replay harnesses.
• Custom or hardware entropy sources — HSMs/TPMs, or a host that already owns a
  vetted CSPRNG and wants a single entropy source.
• no_std/embedded targets where getrandom has no backend and entropy must be
  provided by the caller.

What changed

New *_with_rng methods, each taking rng: &mut impl CryptoRng:

• Account::new_with_rng
• Account::generate_one_time_keys_with_rng
• Account::generate_fallback_key_with_rng
• Account::create_outbound_session_with_rng
• Session::encrypt_with_rng — consumes RNG bytes only when the message triggers a
  Diffie-Hellman ratchet advance; a pure symmetric-chain message draws nothing.
• Curve25519SecretKey::new_with_rng, Curve25519Keypair::new_with_rng
• Ed25519Keypair::new_with_rng, Ed25519SecretKey::new_with_rng

The rng is threaded through internal helpers (Ratchet, RemoteRootKey::advance,
DoubleRatchet, OneTimeKeys, FallbackKeys) via parallel _with_rng methods.
decrypt/create_inbound_session are unchanged — they mint no keys.

Non-breaking / additive

Every existing method keeps its exact body and behaviour, still backed by rand::rng().
The diff is purely additive (the only two removed lines augment two use statements). No
public signature changes; no existing test changed. The audited OsRng code paths are
byte-identical.

Security note (the footgun)

An injected RNG is a sharp tool: the security of every key rests entirely on the quality
of the supplied generator. A low-entropy, predictable, or reused RNG produces
predictable or repeated secret keys, and reusing an RNG state across two genuinely-new
Diffie-Hellman ratchet steps collapses the ephemeral keys and destroys forward secrecy /
post-compromise security. Each new public method documents this explicitly and directs
callers to pass a cryptographically secure generator seeded with fresh entropy per
genuinely-new operation. The default OsRng-backed methods remain the right choice for
almost all users.

Tests

tests/with_rng.rs covers: determinism (identical RNG ⇒ byte-identical keys/session-id/
ciphertext across all seams); distinct RNG ⇒ distinct keys; interop (a _with_rng-built
session round-trips with a default OsRng-built peer); and the lazy DH-ratchet advance
seam (reproducible under the same RNG, fresh ephemeral under a different one). All
existing tests pass; cargo clippy --all-targets --all-features is clean.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 94.81865% with 10 lines in your changes missing coverage. Please review.
✅ Project coverage is 91.14%. Comparing base (6b38b2c) to head (3a1bcb7).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/olm/account/one_time_keys.rs	87.50%	1 Missing and 2 partials ⚠️
src/olm/session/double_ratchet.rs	92.85%	0 Missing and 3 partials ⚠️
src/olm/account/fallback_keys.rs	93.33%	0 Missing and 1 partial ⚠️
src/olm/account/mod.rs	97.67%	0 Missing and 1 partial ⚠️
src/olm/session/mod.rs	96.29%	1 Missing ⚠️
src/olm/session/root_key.rs	93.33%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #379      +/-   ##
==========================================
+ Coverage   90.93%   91.14%   +0.20%     
==========================================
  Files          34       34              
  Lines        4955     5148     +193     
  Branches     4955     5148     +193     
==========================================
+ Hits         4506     4692     +186     
+ Misses        286      285       -1     
- Partials      163      171       +8     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

[codspeed-hq]

Merging this PR will not alter performance

✅ 6 untouched benchmarks

---

_{Comparing adapt-toolkit:feature/with-rng-entropy-injection (3a1bcb7) with main (6b38b2c)}

[poljar]

Is this something you are personally invested in? What's your use-case here?

This PR needs some more work:

1. The *_with_rng() functions should be called by the current methods with the RNG set to UnwrapErr(SysRng).
2. The PR should be split out into smaller commits, ideally one added method would be one commit.
3. If this is done for the olm submodule we should do this for the other submodules as well, could be a separate PR.

[shakhvit]

@poljar thanks for the response.

Yes, I wanted to pass my own entropy into your primitives to have full control over the entropy in the system and drop the OS-random dependency. That's actually what let me compile Olm for a bare-metal target (no getrandom backend), which is the main thing I'm after. I'm maintaining a downstream project built on vodozemac, so this isn't a one-off. I'm invested in getting it right and keeping it maintained upstream. Give me a few days to push the reworked version. Thanks for taking the time to review!

[MatMaul]

@shakhvit do you plan to get back at it soon ?

We would need pretty quickly on our side, and I am happy to take over if you don't have the bandwidth for it.

[shakhvit]

@MatMaul I had no time for this recently. Yes, please, you can take it over. Thanks :)