Improve workload allocation safety

[vikranthreddimasu]

• Summary: Allocate integer batch remainders by largest fractional share with deterministic rank tie-breaking.
• Summary: Sanitize invalid scheduler capacity telemetry and normalize extreme finite values without overflow.
• Summary: Validate public sampler capacity helpers before producing workload weights.
• Security changes: None; authentication, authorization, credentials, transport, and deployment behavior are unchanged.
• Usability changes: Batch splits no longer depend on last-rank remainder assignment or node registration order. Malformed capacity inputs now fail with field-specific errors.
• Tests/checks run: 80 focused scheduler, sampler, integration, and convergence tests passed.
• Tests/checks run: make lint passed, including Ruff and mypy across 55 source files.
• Tests/checks run: make test passed with 1028 passed, 3 skipped, and 2 existing deprecation warnings.
• Tests/checks run: pip wheel build passed; git diff --check passed; diff-only secret scan found no matches.
• Commit: f9557a8 Allocate scheduler remainders fairly.
• Commit: 9b353a6 Harden scheduler capacity weights.
• Commit: 8a32f06 Validate sampler capacity inputs.
• Risks or manual review notes: Low-risk scheduling and validation changes. Review integer rounding behavior when global batch size is smaller than live node count.
• Merge status: Merged into master after all 11 GitHub checks passed and branch remained conflict-free.