fix(llm): a slow model is not a hung connection (SSE stream timeout)

[postgresql007]

Found while standing up the LLM query harness against the DGX endpoints.

A slow model is not a hung connection

The OpenAI-compatible provider set http.Client.Timeout: 5 * time.Minute, with a
comment calling it "a backstop for hung connections". It is not one.
Client.Timeout is a deadline on the whole exchange including the body, and
this is an SSE client — it cannot tell a dead connection from a healthy stream
still delivering tokens.

Measured against the live endpoints:

model	one ordinary question, max_tokens=4096	stream healthy?
deepseek-v4-flash-0731	265 s	yes — continuous bytes, clean data: [DONE]
qwen3.8-flash-next	532 s	yes — same

Both were killed at 300 s and reported as:

openai: read stream: context deadline exceeded
  (Client.Timeout or context cancellation while reading body)

which reads as a network fault rather than our own client hanging up on a
working model. Every reasoning model is affected — and since reasoning
models are exactly what operators point this at, the helper was unusable
against them for any non-trivial question.

The fix

Bound the phases where silence genuinely means a fault, and leave generation
to the caller's ctx:

• connect + TLS handshake — 30 s
• response headers — 5 min (a busy inference server queues before generating)
• gap between bytes mid-stream — 2 min, via a stallReader that resets on
  every Read

That last one is the condition the old timeout was reaching for: a server that
stops sending without closing. A stalled stream now says so, instead of
surfacing a bare context error.

Verification

The query that died at 300 s now returns rc=0 in 265 s, with a real
answer. Provider and internal/llm/... unit tests pass.

[postgresql007]

Superseded by #63.

Both commits on this branch are contained in integration/v1.5.0, together with the #56 systemd work and the dependency bumps. #63 also adds documentation for PG_HARDSTORAGE_LLM_HOT_HELP_BYTES — TestOperatorEnvVarsAreDocumented failed on the integration branch because the new env var was read by production code and appeared nowhere in docs/, which is exactly what that guard is for.

Consolidating means one review and one release. The branch fix/llm-streaming-timeout is left in place; delete it whenever you like.