fix: make Wildfire result paths cross-platform

[carlosfab]

Context

While reproducing the Wildfire MANUAL baseline on macOS from upstream commit
d648f128, I ran one step of Cut_Trees_Sparse_small with seed 483 and three
neutral (0) firefighter actions. The simulation itself reached
TEST COMPLETE, produced a score of 6, and made no API calls, but automatic
post-processing failed with FileNotFoundError.

The Python entrypoint had written data.csv below the literal path
results\logs\MANUAL, while Unity had written 504 frames below
results/logs/MANUAL. A backslash is a path separator on Windows but an ordinary
filename character on POSIX systems, so the CSV and frames ended up in different
directory trees. Rendering succeeded only after manually consolidating the
artifacts under the POSIX tree and invoking the existing renderer directly.

The same hardcoded separator pattern was present in all seven Wildfire
entrypoints and in six agent chat-log implementations.

Changes

• centralize result and agent chat path construction with pathlib while
  preserving string paths at the call boundary;
• use the shared helpers in CAMON, COELA, Embodied, HMAS_2, MANUAL, and both
  algorithm templates;
• make the standalone renderer example platform-independent;
• add regression coverage for POSIX paths, Windows paths, nested agent chat
  paths, and co-location of result and chat artifacts.

This does not change algorithm behavior, prompts, models, dependencies, result
names, seeds, or timestamps.

Verification

• python -m unittest crew_algorithms.wildfire_alg.test_paths -v — 4 tests pass
  under Python 3.10.11;
• Black check passes for the new modules;
• Flake8 passes for the new modules;
• all affected Python modules compile with DeprecationWarning treated as an
  error;
• git diff --check passes;
• no hardcoded results\logs or Agent_N\chats.txt paths remain in the affected
  runtime code.

I also repeated the original one-step smoke test against the macOS Unity build
with ML-Agents 2.3.0-exp.3 / communication protocol 1.5.0. It completed with
exit code 0 and produced one coherent result tree containing:

• data.csv with score 6 and zero API calls/input tokens/output tokens;
• 504 PNG frames across the 8 expected streams;
• a 63-frame MPEG-4 render at 600 x 1200, 20 fps, and 3.15 seconds;
• 506 files in total, with no filename or directory containing a literal
  backslash.

The generated CSV has the same SHA-256 as the pre-fix reference run
(2fe85b5ffbe80cca82c076f8af3964c28ad23fdf2de164d7ca4439d9d7c67712), while
the process now completes without the manual artifact relocation previously
needed for rendering.