Allow block fleets to grow for smaller jobs

[SiluPanda]

Summary

• query new VM capacity with fleet-sized resources, then match each candidate against the job's proportional block requirements
• persist the full host offer and selected busy-block count while exposing a fractional offer to the job, so remaining blocks can be reused
• preserve CPU architecture and configurable disk constraints, and keep unsupported container or wrapper-emitted offers out of the shared-host path before offer limits are applied

SSH block fleets retain their existing requirement-combination behavior, and VM provisioning still receives the job's runtime requirements.

Fixes #4188.

Validation

$ PYTHONPATH=src uv run --group test pytest -q <seven focused block-fleet tests>
7 passed, 4 skipped in 0.30s

$ PYTHONPATH=src uv run --group test pytest -q src/tests/_internal/server/services/runs/test_plan.py::TestGetBackendOffersInFleet
5 passed, 4 skipped in 0.21s

$ PYTHONPATH=src uv run ruff check <changed Python files>
All checks passed!

$ PYTHONPATH=src uv run ruff format --check <changed Python files>
8 files already formatted

$ git diff --check origin/master...HEAD
# no output

AI-assisted implementation; I reviewed the final diff, independently peer-reviewed the edge cases, and ran the checks above locally.