feat: configure Ray GPU scheduling fractions

[skirdey]

Summary

• expose the hard-coded Ray training and rollout actor GPU claims as CLI options
• preserve the existing 0.4 / 0.2 defaults
• validate each claim and reject impossible colocated sums above one GPU
• document that these are scheduling claims, not CUDA utilization caps

Motivation

The actor resource claims are currently hard-coded in two different modules. This makes unusual single-GPU and colocated layouts difficult to tune and can leave a job waiting indefinitely if downstream patches create incompatible claims. This addresses the configurability request in #379 while keeping current behavior unchanged by default.

Test plan

• PYTHONPATH=$PWD uv run --no-project --with pytest --with pyyaml --with omegaconf pytest -q tests/test_megatron_argument_validation.py — 24 passed
• PYTHONPATH=$PWD /tmp/slime-ray312/bin/python -m pytest -q tests/test_placement_group.py — 12 passed
• python3 -m py_compile ...
• git diff --check