fix(policy): prevent policy callback deadlocks and teardown leaks

[rvatkar]

Fixes DCGM-3746 and DCGM-6904.

Both issues came from the same root cause: ViolationRegistration could block while DCGM held internal locks.

Summary

• make ViolationRegistration non-blocking by dropping on full per-condition queues and tracking drops with droppedPolicyViolations
• sample drop logs at powers of two so overload does not turn the callback path into log I/O
• start fan-in before dcgmPolicyRegister_v2 and make fan-in sends cancellation-aware so registration and teardown cannot hang on a full user-facing channel
• increase the per-condition buffer from 1 to 16 to absorb short bursts

Why policyChannelBuffer = 16

• DCGM_MAX_XID_INFO = 10 (from dcgm_structs.h) is the upstream design ceiling for in-flight XID events per callback, so the buffer must be >= 10 to absorb a worst-case synchronous burst.
• 16 is the next power of two above 10, giving small headroom for scheduler jitter without straying far from what DCGM itself considers the realistic ceiling.
• Memory cost is negligible: 7 conditions x 16 slots x ~64 bytes per PolicyViolation ~= 7 KB total. Sized larger (e.g. 64) would be hard to defend in review for no measurable benefit.
• If droppedPolicyViolations ever becomes non-zero in production, that is the signal to revisit the size -- not a guess.

Tests

Added deterministic regression coverage for:

• bounded callback latency on full per-condition channels
• exact drop accounting
• recovery after a dropped violation
• fan-in cancellation while blocked on output
• registerPolicy pre-registration failure cleanup
• normal lifecycle cleanup on caller cancellation
• registration-failure cleanup for both registerPolicy and registerPolicyOnly

[rvatkar]

Closing this PR, as #119 handles it.