feat: add DFlash for Windows

[aminya]

This adds support for building Dflash on Windows

cmake -B build -S . -DCMAKE_CUDA_ARCHITECTURES=86 -DCMAKE_BUILD_TYPE=Release
cmake --build build --parallel

Tested with Qwen 27B on RTX3090 compiled with MSVC + Cuda. and got 93.53 tok/s

### 2. Recursive
[timing] per-step averages over 35 steps (ms):
  draft_build    0.39
  draft_copyfeat 0.10
  draft_set      0.06
  draft_compute  9.40
  draft_logits   7.79
  snapshot_ssm   0.00
  verify_build   2.76
  verify_set     0.84
  verify_compute 45.92
  verify_logits  0.00
  accept         8.58
  restore_ssm    0.00
  replay_build   0.00
  replay_set     0.00
  replay_compute 0.00
  replay_logits  0.00
  ----- sum     75.85

[dflash] generated 256 tokens in 2.737 s  →  93.53 tok/s
[dflash] 35 draft steps, accepted=256/560 (45.7% per step), avg commit/step=7.31
[dflash] output tail: 11 292 283 292 11 264 478 292 198 262 460 292 198 71093 271 13962 220 17 13 82930 
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 24575 MiB):
  Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes, VRAM: 24575 MiB

[run] generated 256 tokens

Also, tested with Qwopus on RTX3090 compiled with MSVC + Cuda and got 67.24 tok/s. Using the Qwen 27B kernel, so the acceptance was lower due to mismatch

[run] prompt 14 tokens, streaming up to 256 tokens…
[cfg] seq_verify=0 fast_rollback=1 ddtree=1 budget=22 temp=1.00 chain_seed=1
[target] target loaded: 851 tensors on GPU 14.73 GiB, tok_embd 682 MiB CPU-only (q4_K)

[timing] per-step averages over 48 steps (ms):
  draft_build    0.57
  draft_copyfeat 0.12
  draft_set      0.06
  draft_compute  9.64
  draft_logits   8.36
  snapshot_ssm   0.00
  verify_build   3.03
  verify_set     1.08
  verify_compute 46.26
  verify_logits  0.00
  accept         8.91
  restore_ssm    0.00
  replay_build   0.00
  replay_set     0.00
  replay_compute 0.00
  replay_logits  0.00
  ----- sum     78.03

[dflash] generated 256 tokens in 3.881 s  →  65.96 tok/s
[dflash] 48 draft steps, accepted=256/768 (33.3% per step), avg commit/step=5.33
[dflash] output tail: 11 307 478 220 16 1590 198 1734 264 11 292 283 292 11 264 478 292 198 260 460 
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 24575 MiB):
  Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes, VRAM: 24575 MiB

[run] generated 256 tokens