Support APIC serialization for deterministic CUDA launches

[shi-eric]

Description

Follow-up to #1443 and 24ee010.

Deterministic CUDA kernels currently cannot be serialized through APIC. When a deterministic CUDA kernel is launched inside wp.ScopedCapture(..., apic=True), Warp raises a clear RuntimeError explaining that APIC serialization is not currently supported for deterministic CUDA kernels.

This limitation is intentional in the initial deterministic execution work. Deterministic CUDA launch is not a single kernel launch with extra arguments; it may allocate and retain scratch buffers, reset counters and overflow state, launch phase-0 and phase-1 kernels, run native sort/reduce or counter scan/writeback helpers, preserve buffers for CUDA graph replay, and sometimes perform host-side overflow checks.

Desired behavior

APIC capture should either:

• serialize deterministic CUDA launches using a defined deterministic APIC record format, or
• expand deterministic launches into APIC-recordable primitives that include the hidden deterministic buffers and native helper calls.

Acceptance criteria

• Deterministic accumulation kernels can be captured with apic=True, saved, loaded, and replayed correctly.
• Deterministic consumed-return counter kernels can be captured with apic=True, saved, loaded, and replayed correctly.
• Hidden deterministic buffers and native helper calls have well-defined lifetime and replay behavior.
• Unsupported deterministic/APIC combinations continue to fail with clear errors.
• Tests cover APIC capture rejection or support for both scatter/reduce and counter deterministic paths.