perf: lazy constant resolution in ErgoTree evaluation

[mwaddip]

Resolve ConstPlaceholder from the eval context on demand instead of deep-cloning the AST via proposition() each reduce_to_crypto call. Mainnet-sync profiling in ergo-node-rust showed the per-input clone-and-substitute was 8-12% of total CPU.

Adds root_expr() / constants() accessors and Context::with_constants(). Deserialize trees fall back to proposition() (incompatible with lazy resolution). proposition() itself unchanged.

Beware of conflicts with #854, see comments below

[mwaddip]

Rebase guidance update (re: #876 — "JIT Costing: The Definitive PR")

Update: the JIT-costing work was re-sliced into #876 (the "Definitive" PR), which
supersedes #854. The reconciliation below is unchanged in substance — #876 carries the
same proposition_for_cost_eval + ConstantPlaceholder::resolved machinery and already
charges ConstPlaceholder at JitCost(1).

#876 and this PR overlap in reduce_to_crypto / ConstPlaceholder: #876 resolves segregated
placeholders via ConstantPlaceholder::resolved (populated by proposition_for_cost_eval);
this PR resolves them via Context.constants. The ergo-node-integration branch carries both
layered — see 9094f693 for the worked reconciliation.

If #876 merges first → this PR rebases:

• Add constants: Option<&'ctx [Constant]> to Context alongside the jit_cost fields.
• reduce_to_crypto: keep the deserialize path; refactor the common path to tree.root_expr()
  • with_constants(). Preserve cumulative-cost tracking by syncing the cloned Cell back after
    inner returns (ctx.jit_cost.set(ctx_with_c.jit_cost_value())).
• Expr::ConstPlaceholder eval arm: resolve via ctx.constants (cost stays JitCost(1)).
• JIT Costing: The Definitive PR #876's ConstantPlaceholder::resolved field becomes redundant — drop it (or keep as no-op).
• trivial_reduce: resolve segregated placeholders via ctx.constants on the common path.

If this PR merges first → #876 rebases: add jit_cost / jit_cost_limit alongside
constants; add the cost_before snapshot + post-inner sync on the common path; drop the
resolved field; switch the ConstPlaceholder arm to resolve via ctx.constants (cost already
1); retune the placeholder/limit tests.

Net: whichever merges first, the ConstPlaceholder arm ends up charging 1 JitCost and reading
from ctx.constants. The rest is plumbing.