fix(scheduler): keep tenant queue caps opt-in

[DavidBellamy]

Problem

SMG prod currently defaults fair_share.max_queued_requests_per_tenant to 64. Existing fair-share configurations omit the new field, so adopting the current prod wheel would silently activate a restrictive queue cap even though that behavior has not been separately reviewed or configured.

Required safety behavior

• Omitting max_queued_requests_per_tenant preserves the pre-change occupancy behavior.
• None resolves to the scheduler partition's existing shared queue bound.
• An explicit positive value remains available for future, separately reviewed queue policy.
• Zero remains invalid.
• No production image or runtime activation accompanies this source fix.

Acceptance

• Backward-compatible YAML parsing test proves omission is inert.
• Explicit value and zero-validation tests remain covered.
• Queue tests prove the omitted setting does not add a tenant-specific ceiling below the shared budget.