fix(openresponses): support Codex WebSocket warm-up

[aim9sour]

Description

This follows up on #8644 and #8676 and fixes remaining Responses WebSocket compatibility gaps exposed by Codex:

• supports generate:false warm-up without inference and keeps store:false responses in a connection-local store
• resolves multi-hop previous_response_id chains across connection-local and global stores
• keeps native Responses tools such as web_search and namespace from being coerced into function tools
• preserves the requested store flag in WebSocket response resources
• atomically releases the one-in-flight guard at the terminal-event boundary so immediate sequential requests do not get a false response is already in progress error

Regression coverage includes connection isolation, warm-up continuation, mixed-store ancestry, native tool filtering, and immediate sequential turns.

Notes for Reviewers

Verified on the final commit:

• go test ./core/http/endpoints/openresponses ./core/schema --count=1
• go test -race ./core/http/endpoints/openresponses --count=1
• go vet ./core/http/endpoints/openresponses ./core/schema
• Codex WebSocket warm-up E2E repeated four total runs (--repeat=3)
• all 8 WebSocket Responses E2E specs
• make lint with the CI toolchain (Go 1.26.1 + golangci-lint 2.11.4): 0 issues
• make build with Go 1.26.1

A broader make test was also attempted in the earlier reproduction worktree. It reached unrelated container-backed tests and failed because this environment has no Docker/rootless Docker provider (docker is not installed and no Docker socket/runtime is present). The targeted, race, lint, E2E, vet, and build checks above are green.

Per the repository AI contribution policy, the commit includes an Assisted-by trailer and is DCO-signed by the author.

[Signed commits]

• Yes, I signed my commits.
• Documentation updated or not applicable

[localai-org-maint-bot]

The WebSocket changes look well covered, but this PR changes user-facing Responses API behavior without updating the corresponding guide. LocalAI's docs-with-code rule requires that update in the same change.

Please add a short WebSocket Responses section to docs/content/features/text-generation.md that documents:

• connecting to /v1/responses over WebSocket and waiting for a terminal event before sending the next response.create;
• generate: false returning response.created and response.completed without inference, and using its ID as previous_response_id;
• store: false response IDs being scoped to the current WebSocket connection; and
• multi-hop previous_response_id chains replaying the complete conversation.

I did not push this minor docs fix because a bot-authored commit would fail DCO, and I cannot preserve your sign-off on content you did not commit.