feat: add step-scoped H2D .to() timing

[ppraneth]

What this PR does

Closes #82.

Adds instrumentation for host-to-device transfers so TraceML can measure how long tensor.to(cuda_device) takes during training steps. This is a meaningful signal in data-heavy pipelines where the GPU can sit idle waiting for DMA to finish.

How it works

Automatic mode patches torch.Tensor.to() once at traceml.init() time. The patch is gated by a thread-local flag that is only raised inside trace_step(), so model initialization, checkpoint loading, and any other setup transfers are completely ignored. Only CUDA-targeted calls are timed -- dtype-only casts and CPU-to-CPU copies pass through with zero overhead.

Manual and selective modes get an explicit traceml.wrap_h2d(x) wrapper. It returns a thin proxy around the tensor or batch object so the next .to(device) call is timed. No context manager, no broader API change -- just wrap and go:

x = traceml.wrap_h2d(x)
x = x.to(device)

Both paths emit CUDA events (same approach as forward/backward timing) and feed into the existing StepTimeSampler pipeline, so the measurement lands in the step_time_samples SQLite table automatically alongside all other training signals.

Files changed

• src/traceml/utils/patches/h2d_auto_timer_patch.py -- new patch module with TLS gating, CUDA target detection, and the h2d_auto_timer context manager
• src/traceml/instrumentation.py -- h2d_auto_timer nested inside trace_step alongside forward_auto_timer and backward_auto_timer
• src/traceml/initialization.py -- patch_h2d field added to TraceMLInitConfig, wired into init() and start()
• src/traceml/wrappers.py -- _WrappedH2D proxy class and wrap_h2d() function with duplicate instrumentation guard
• src/traceml/api.py -- wrap_h2d exposed at the public API layer; init() and start() accept patch_h2d
• src/traceml/__init__.py -- wrap_h2d added to top-level exports
• tests/test_h2d_timing.py -- 30 tests covering TLS gating, CUDA target detection, step scoping, manual wrapper, duplicate guards, and init config

What is not in this PR

No rendering or UI changes. The _event_bucket summary function does not yet map h2d_time to a display bucket -- that is a follow-up consistent with what the issue specifies.

Testing

All 30 tests pass without a GPU. Timing falls back to CPU wall-clock when CUDA is unavailable, so the full test suite runs in any environment.

python -m pytest tests/test_h2d_timing.py -v
# 30 passed in 1.81s