fix: tolerate null response content joins

[giulio-leone]

Summary

• tolerate TextContentItem(text=None) when flattening interleaved content
• tolerate None entries when joining ChatCompletionResult.content_text
• add focused regressions for both null-content join paths

Why

Current main still reproduces the null-content crash from #4996:

1. interleaved_content_as_str([TextContentItem.model_construct(text=None)])
2. ChatCompletionResult(..., content=[None, "tool result"], ...).content_text

Both currently raise:

TypeError: sequence item 0: expected str instance, NoneType found

The earlier closed attempt in #5029 fixed the same underlying issue, but it targeted the older meta_reference responses layout. After the current builtin/responses refactor, the same bug is still present on main under the new path.

Validation

• .venv/bin/ruff check src/llama_stack/providers/utils/inference/prompt_adapter.py src/llama_stack/providers/inline/agents/builtin/responses/types.py tests/unit/providers/utils/inference/test_prompt_adapter_none_safety.py tests/unit/providers/inline/agents/builtin/responses/test_types.py
• .venv/bin/python -m pytest tests/unit/providers/utils/inference/test_prompt_adapter_none_safety.py tests/unit/providers/inline/agents/builtin/responses/test_types.py tests/unit/providers/inline/agents/builtin/responses/test_streaming.py -q
• exact source-tree proof on clean origin/main vs this branch showing main still raises TypeError in both paths while this branch returns "" / "tool result"

Closes #4996
Supersedes #5029

⚠️ This reopens #5244 which was accidentally closed due to fork deletion.

[mergify]

This pull request has merge conflicts that must be resolved before it can be merged. @giulio-leone please rebase it. https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[cdoern]

This PR is out of date, please rebase.

[giulio-leone]

@cdoern rebased this PR onto current main and pushed the updated branch.

Notes:

• the upstream agents -> responses rename moved one test import path during the rebase; the fix itself stayed the same
• targeted regression validation passed after the rebase with:
  • uv run --python 3.12 --group dev --group unit pytest tests/unit/providers/inline/responses/builtin/responses/test_types.py tests/unit/providers/utils/inference/test_prompt_adapter_none_safety.py

This should clear the out-of-date/conflict blocker from the earlier review.

[giulio-leone]

@cdoern quick follow-up: the earlier out-of-date/rebase blocker should now be cleared on the current head.

The PR was rebased onto current main, the targeted regression validation was rerun after the rebase, and the effective PR check state is now 72 successful, 0 failing, 2 skipped (with only non-blocking cancelled entries in the matrix history).

If you have a moment, I would appreciate a fresh look to confirm there are no remaining blockers. Thanks!