CUDA: better coalesce data-access for contiguous concat

[ORippler]

Overview

Stumbled upon this when looking at an nsight trace for a hybrid mamba model.
Nsight compute revealed uncoalesced data-access (especially stores) for typical LLM token-gen workloads in addition to launching a lot of CTAs. This PR addresses these two points, leading to 1-3% E2E perf gain.

I did not find perf tests in test-backend-ops for this, so I didn't have any other workloads to verify against perf regression (PP hits the non-cont concat_f32_non_cont kernel).

Additional information

• ./scripts/compare-llama-bench.py -b 15fa3c4 -c osimons/CUDA_faster_cont_concat --tool llama-bench -i llama-bench.sqlite

Model	Test	t/s master	t/s osimons/CUDA_faster_cont_concat	Speedup
nemotron_h_moe 31B.A3.5B NVFP4	tg128	188.30	192.70	1.02
nemotron_h_moe 31B.A3.5B NVFP4	tg128@d32768	173.95	175.01	1.01
nemotron_h_moe 31B.A3.5B Q4_K_M	tg128	211.40	218.11	1.03
nemotron_h_moe 31B.A3.5B Q4_K_M	tg128@d32768	199.50	204.58	1.03
qwen35 27B Q4_K_M	tg128	39.42	40.08	1.02
qwen35 27B Q4_K_M	tg128@d32768	35.65	36.21	1.02
qwen35moe 35B.A3B Q4_K_M	tg128	164.40	169.44	1.03
qwen35moe 35B.A3B Q4_K_M	tg128@d32768	143.53	147.15	1.03

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, scoped the task for it to solve and reviewed, verified + cleaned up the code afterwards