Add configurable GPU launch wrappers

[akrivx]

Summary

Introduce free-function GPU launch wrappers to centralise direct kernel launches and indirect GPU operations (library calls / composite ops), enabling optional instrumentation (timing, NVTX, error checking) configured statically at compile time.

Design doc: docs/design/gpu-launch-wrappers.md

Motivation

We want a single convention for submitting GPU work that:

• keeps call sites simple,
• avoids runtime overhead when instrumentation is disabled,
• enables optional profiling/logging/validation,
• supports both direct kernel launches and indirect "GPU ops".

Scope

In scope

• Free functions: launch_kernel, launch_op
• Compile-time policy selection
• Optional instrumentation compiled out when disabled
• Nested launch wrapper calls are not allowed:
  • Calling launch_kernel or launch_op from within an active wrapper region is invalid.
  • Must be detected (at least in debug builds).

Out of scope

• Mandatory enforcement (raw <<<...>>> still allowed)
• Intercepting internal library kernel launches
• Runtime configuration
• CUDA Graph capture integration (for now)

Acceptance Criteria

• launch_kernel() wraps kernel launches
• launch_op() wraps indirect GPU work (e.g. cuBLAS/CUB calls or composite ops)
• Non-nested invariant enforced (debug assert / optional hard error)
• Behaviour configurable at compile time via policy type
• Zero runtime overhead when features disabled
• No launcher objects required
• Semantics match design doc