v0.12.0 tracking issue

[coreylowman]

This is going to be a smaller breaking release focused around f16 and concat/stack. This is to support the development of transformer models.

• Add f16 dtype support for tensors #423
• Concat along any dimension #731
• Dilation for Conv2d and MaxPool2d ?  #693
• Add Conv2D groups #589
• Support for conv1d #219
• Add AMP dtype #424

Lower prio, but nice to have:

• Change dropout to generate tensor of bools instead of generic dtype #721
• Using no_std on thumbv6m-none-eabi #718

[coreylowman]

Note: this is currently blocked on a new release of the half-rs crate.

[ViliamVadocz]

New version of half has been released. Is this no longer blocked?

[coreylowman]

Yep, and the 0.9.11 version of cudarc uses that, so once I update the conv transpose for dilation/groups I'll push the release. Probably won't add AMP for now