Remove ftz

Avoid ci errors

Note that crossbeam is an indirect dependency and they are still to release a version that [passes](crossbeam-rs/crossbeam#996) on this miri check.

So to temporarily get this out of the way, you can patch crossbeam on the workspace `dfdx/Cargo.toml`.

Thread main finished before other threads were still active. The only
thing related to threads were the gemm usage of rayon, and disabling
rayon usage for matmul seems to be sufficient to fix this.

- Moved the impl of `Cache::try_empty_cache()` to `TensorCache::clear()`.
  - This can be invoked both by `Cache::try_empty_cache()` and by `drop(TensorCache)`.
- Moved the device cache ptr deallocation to `BytesPtr`, `CudaBytesPtr` (newtype over `CUdeviceptr`) and `Buffer`.
  - This is abstracted by the `CachePtr` trait.
  - Can be called by `TensorCache::clear()`.
  - This method may require some "extra" device information, such as in the cuda case. That information is held by `TensorCache`.