0.3.0

Faster CUDA kernel and dispatch to other dtypes

• Add a script to benchmark kernels between torchdtw version: benchmark/src/dtw_benchmark/compare_release.py
• Dispatch to all kinds of dtypes for distances and sx in CPU and CUDA kernels.
• Faster CUDA kernel:

[---------------- dtw / cpu -----------------]
                 |  released 0.2.0  |  current
1 threads: -----------------------------------
      64x64      |         9.2      |      9.4
      128x128    |        23.9      |     24.4
      256x256    |        85.6      |     86.1
      512x512    |       327.1      |    327.8
      1024x1024  |      1291.0      |   1295.6

Times are in microseconds (us).

[-------------------- dtw_batch / cpu --------------------]
                              |  released 0.2.0  |  current
1 threads: ------------------------------------------------
      n=4x4, s=128x128, sym   |        72.8      |     72.8
      n=4x4, s=128x128, asym  |       155.6      |    154.9
      n=8x8, s=256x256, sym   |       665.9      |    663.5
      n=4x8, s=256x512, asym  |      1899.2      |   1901.0

Times are in microseconds (us).

[---------------- dtw / cuda ----------------]
                 |  released 0.2.0  |  current
1 threads: -----------------------------------
      64x64      |        96.7      |     81.3
      128x128    |       149.4      |    123.6
      256x256    |       288.0      |    228.2
      512x512    |       745.2      |    560.8
      1024x1024  |      2581.7      |   1159.4

Times are in microseconds (us).

[-------------------- dtw_batch / cuda -------------------]
                              |  released 0.2.0  |  current
1 threads: ------------------------------------------------
      n=4x4, s=128x128, sym   |      100.0       |    81.3
      n=4x4, s=128x128, asym  |      104.2       |    84.1
      n=8x8, s=256x256, sym   |      305.3       |   153.3
      n=4x8, s=256x512, asym  |      514.8       |   353.0