fix: deepseek v4 inter-leaved tool and reasoning parser - #8901

Overview:

Fixes DeepSeek V4 tool-continuation reasoning parsing without regressing Kimi K2.5. In the failing DSv4 flow, the request ends with a role=tool message and the DSv4 formatter has already seeded <think> into the next assistant prompt. The model can then stream reasoning text followed by </think> without an opening <think> token in the completion. If the postprocessor clears prompt_injected_reasoning for all tool-continuation turns, the closing tag is treated as normal content and leaks to the client.

This PR keeps DSv4 in injected-reasoning mode for that path, while preserving the Kimi K2.5 behavior introduced by 5671d8e374021ca1aeb399fa9729285903ea6645: direct post-tool final answers for kimi_k25 should remain normal content, not be mislabeled as reasoning_content.

Details:

• Removes the broad last_is_tool suppression that caused the DSv4 </think> leak.
• Reintroduces post-tool reasoning suppression specifically for kimi_k25, where the parser is force-reasoning and direct tool follow-up answers must bypass reasoning parsing.
• Adds a DSv4 stream regression test for user -> assistant(tool_call) -> tool -> assistant where the completion starts inside reasoning and emits </think>.
• Adds a Kimi K2.5 regression test for the same conversation shape where the final direct answer must stay in content.
• Keeps tool_choice=required / named guarded from reasoning parsing for guided-decoding JSON outputs.

Where should the reviewer start?

• lib/llm/src/preprocessor.rs — parser-specific post-tool reasoning gate.
• lib/llm/tests/postprocessor_parsing_stream.rs — DSv4 </think> leak regression and Kimi K2.5 non-regression coverage.
• lib/llm/tests/test_reasoning_parser.rs — parser-level DSv4 injected-reasoning tests.

Validation:

Ran with CUDARC_CUDA_VERSION=12090 because local CUDA 13.2 is rejected by cudarc 0.19.3 during build.

• cargo test -p dynamo-llm --no-default-features --test postprocessor_parsing_stream tool_continuation -- --nocapture
• cargo test -p dynamo-llm --no-default-features --test test_reasoning_parser deepseek_v4_post_tool -- --nocapture
• cargo test -p dynamo-llm --no-default-features --test test_reasoning_parser kimi_k25 -- --nocapture
• cargo test -p dynamo-llm --no-default-features --test tool_choice -- --nocapture
• cargo test -p dynamo-parsers _v4 -- --nocapture
• cargo test -p dynamo-parsers tool_calling::xml::kimi_k2_parser::tests -- --nocapture

Related Issues:

• Relates to fix(frontend): fix the rust frontend tool calling with guided decoding and some model parsers fixes #8442 / 5671d8e374021ca1aeb399fa9729285903ea6645