[PERF] Reduce global inference stalls from Tensor.Destroy lock granularity

[tazarov]

Summary

Tensor.Destroy() currently acquires ortCallMu.Lock() (global write lock) in /Users/tazarov/GolandProjects/onnx-purego/ort/tensor.go.

This is correct for safety but stalls unrelated Session.Run() calls across all sessions while any tensor is being released.

Context

• File: /Users/tazarov/GolandProjects/onnx-purego/ort/tensor.go
• Current behavior is explicitly documented and TODO-marked in code.
• PR context: feat: implement pure-go tensor/session inference MVP (issues #3, #4) #37

Why this matters

In multi-session inference pipelines, batch/tensor rotation can cause avoidable global contention.

Proposed direction

1. Design finer-grained lifetime tracking so releasing one tensor does not block unrelated in-flight runs.
2. Keep environment teardown safety guarantees intact.
3. Preserve pure-Go/no-CGO constraints.

Potential approaches:

• Per-handle active-use tracking (reference counters), or
• Session/value dependency tracking for targeted waits.

Acceptance criteria

• Tensor release no longer globally blocks unrelated session inference.
• Existing concurrency safety tests continue to pass.
• New regression tests demonstrate reduced cross-session contention without unsafe releases.

Notes

This is a performance/scalability follow-up, not a correctness blocker for MVP.