fix(http): outbound fetch calls have no timeout or redirect control

[BryanFRD]

Problem

Every outbound HTTP call uses bare fetch with no timeout, no redirect policy, and no response-size cap on the network read:

• apiRequest — packages/mcp-core/src/api-client.ts
• exchangeCodeForToken — packages/mcp-core/src/auth/oauth.ts
• fetch_docs — packages/mcp/src/tools/docs.ts (caps the stripped output at MAX_BYTES, but await res.text() reads the whole body into memory first, and follows redirects by default)

In fetch_docs the host comes from a fixed enum, but slug is concatenated unvalidated and redirects are followed silently, so a product host that 30x-redirects offsite would be followed without the host allowlist applying to the final URL.

Why it matters

• A hung or slow upstream (api.ferrlabs.com, a product marketing site, the OAuth exchange) blocks the tool call indefinitely — over stdio that wedges the whole MCP session with no error.
• Following redirects past the vetted host defeats the point of the fetch_docs enum allowlist (mild SSRF surface).
• res.text() with no streamed size guard lets a large/hostile response balloon memory before the post-hoc slice.

Proposed approach

• Add an AbortController timeout (e.g. 15s, configurable) to all fetch calls; map abort to a clear tool error.
• In fetch_docs, set redirect: 'manual' (or re-check the final URL host against the allowlist) and reject the slug if it contains ://, .., or a leading //.
• Cap the read for fetch_docs by reading the stream incrementally and stopping at MAX_BYTES.

Acceptance criteria

• A slow upstream produces a timeout error instead of hanging.
• fetch_docs does not follow a redirect to a non-allowlisted host.
• Tests cover timeout and redirect-rejection paths.