fix: preserve Responses reasoning stream lifecycle

[chrisnestrud]

Summary

Fixes the Anthropic-to-Responses reasoning stream lifecycle used by Codex before tool-call continuations.

• Initialize the reasoning summary_text part before emitting deltas.
• Emit reasoning text/part/item completion events with the provider signature.
• Carry the Anthropic message ID into the Responses lifecycle.
• Preserve signature_delta rather than treating it as a generic thinking delta.
• Add a mocked E2E replay: Anthropic thinking + tool use → Responses output → Anthropic tool-result continuation.

Fixes #102
Refs #63
Refs #101

Validation

CGO_ENABLED=0 go test -p 1 ./...
go test -tags=e2e ./internal/e2e -run "^TestOpenAIResponseAnthropicE2E_ReasoningToolReplay$"

The complete tagged E2E suite has unrelated existing failures tracked in #101.

[chrisnestrud]

Cross-reference: while diagnosing the same reported symptom I found a second,
independent defect in the replay path.

prependCachedThinking on the adapter path only inserted a thinking block for
assistant turns containing tool_use, so a conversation whose earlier assistant
turn was plain text reached the provider bare and was rejected with
The content[].thinking in the thinking mode must be passed back to the API.
Every retry re-sent the same body, so the session died rather than recovering.
Replaying a captured failing request against the provider confirmed it: the body
returns 400 as-is, and 200 once the plain assistant turn gains a thinking block
(empty or filled).

Filed as #110, fix in #111, based on main independently of this PR.

This PR is still needed for the reasoning summary lifecycle; the replay gap only
shows up once a session contains a mix of tool-call and text-only assistant
turns, which is why a short probe passes while longer sessions fail.