Fix eval_duration None TypeError in format_utils.py (#162)

[AmirF194]

Root cause: ollama_chat_to_openai_v1_chat_completion and the streaming final-chunk
handler in format_utils.py read the four duration fields with
ollama_response.get("eval_duration", 0) (and the three siblings). dict.get's
default only applies when the key is absent, not when it is present with value
None. A short or aborted generation, or a prompt-cache hit, returns these keys
present and None, so eval_duration_ns > 0 raises
TypeError: '>' not supported between instances of 'NoneType' and 'int' and the
worker dies, exactly as reported in #162.

Fix: get(key) or 0 in place of get(key, 0) for all four duration fields, in
both the non-streaming function and the streaming path's final chunk, so a
None value falls back to 0 the same way a missing key already does.

Note: PR #164 fixed the identical crash in server_utils.py, but never touched
format_utils.py, so this specific path stayed broken; the two fixes are
independent.

Verification:

• Added test_format_utils.py (loaded by file path, since rkllama/api/__init__.py
  pulls in the RKNN NPU runtime at import time, which is unavailable off Rockchip
  hardware): 3 cases with a duration key set to None, on both the chat and the
  streaming path, fail on main with the exact reported TypeError and pass on
  this branch; a 4th case confirms real durations still convert correctly.
• Ran in a clean python:3.12-slim container with only format_utils.py's own
  dependencies installed (flask, cv2, numpy, pillow, requests, pydantic); no NPU
  hardware needed since the bug is in response formatting, not inference.
• Not verified: the actual RK3588 inference path that produces a None duration
  end-to-end (no NPU hardware in this environment); the repro constructs that
  input directly, matching the shapes described in TypeError: '>' not supported between 'NoneType' and 'int' in format_utils.py when eval_duration is None → worker dies #162 and by the merged FIx issue #162 with eval_duration None #164.

Fixes #162

[AmirF194]

Checking in on this one week later, no rush at all. It fixes a real crash: eval_duration (and the other duration fields) can come back None on a short or cached generation, and the current get(key, 0) default only covers a missing key, not a None value, so the request handler throws. Happy to adjust anything if the fix shape needs changes.