[NV] Allow ext dist coeffs tensors to live on CPU

[frtrudel]

!! KEEPING AS DRAFT UNTIL PREREQUISITES CHANGES ARE MERGED !!

Allows external distortion coefficients tensors to live on the CPU and avoid potential CPU-GPU sync(s).

Metric	Baseline	Fix	Delta
Training step avg	134.4 ms	134.1 ms	-0.3 ms (-0.2%)
Training step median	119.6 ms	120.0 ms	+0.4 ms (+0.3%)
cudaStreamSync total	82.6 s	83.9 s	+1.3 s (noise)
DtoH GPU time	121 ms	111 ms	-10 ms
DtoH copy count	19,549	17,353	-2,196
cudaMemcpyAsync CPU time	2,283 ms	1,317 ms	-966 ms
DtoD volume	286,757 MB	286,758 MB	~0

Verdict: The fix correctly removes the targeted DtoH copies (~1 per step per type), but the performance impact is negligible. The removed copies were too small (~2-3 us each) to cause meaningful synchronization stalls.

What the Fix Changed

The fix removes two small DtoH Device→Pageable copies per training step by creating the ext dist coefficient tensor on the CPU/host side instead of on GPU:

DtoH copy size	Baseline (per profile)	Fix (per profile)	Delta
24 bytes	1,126	28	-1,098 (~1/step)
60 bytes	1,102	4	-1,098 (~1/step)
All other sizes	identical	identical	0

Total DtoH Device→Pageable: 5,605 → 3,409 (-2,196 copies, -11 ms GPU time).

The remaining 28 and 4 copies at those sizes are one-time operations (not per-step). The removed copies were 24-byte and 60-byte DtoH transfers at ~2-3 us each. At that size, they complete in a single PCIe transaction and don't create a pipeline stall.

Full Memory Transfer Comparison

Operation	Baseline	Fix	Delta
DtoD	1,053 ms / 62,300 ops / 286,757 MB	1,065 ms / 62,300 ops / 286,758 MB	noise
HtoD	895 ms / 19,352 ops / 10,580 MB	894 ms / 18,620 ops / 10,580 MB	-732 ops (same data)
DtoH	121 ms / 19,549 ops / 119 MB	111 ms / 17,353 ops / 118 MB	-2,196 ops / -10 ms
Memset	67 ms / 46,173 ops	67 ms / 46,173 ops	0

Synchronization Comparison

Metric	Baseline	Fix	Delta
cudaStreamSync calls	26,801	23,873	-2,928
cudaStreamSync total	82.6 s	83.9 s	+1.3 s (noise)
cudaStreamSync avg	3.08 ms	3.51 ms	+0.43 ms
EventSync calls	653,534	667,409	+13,875

The fix has 2,928 fewer sync calls but each is slightly longer on average.

NVTX Training Step Comparison

Phase	Baseline	Fix	Delta
training_step	134.4 ms	134.1 ms	-0.3 ms
training_losses	113.5 ms	114.0 ms	+0.5 ms
forward_gaussians	92.5 ms	93.6 ms	+1.1 ms
render_lidar	77.5 ms	78.7 ms	+1.2 ms
GSplatRenderer/render	16.8 ms	16.8 ms	0 ms
manual_backward	18.6 ms	17.7 ms	-0.9 ms
optimizer_step	1.97 ms	1.95 ms	0 ms

All within run-to-run noise (1-2%).

Top Kernel Comparison

Kernel	Baseline	Fix	Delta
rasterize_bwd	68,442 ms	68,552 ms	+0.2%
rasterize_fwd	34,747 ms	34,832 ms	+0.2%
projection_ut_fused	13,530 ms	13,587 ms	+0.4%
collect_params_2_bwd	7,354 ms	7,315 ms	-0.5%
multi_tensor_apply	7,302 ms	7,350 ms	+0.7%

All kernel times are within noise.