make gpu(x) = gpu_device()(x)

[CarloLucibello]

As a follow-up to #2492, this removes every residual implementation for data movement between devices in favor of relying entirely on MLDataUtils.jl. The 2 PRs combined save us ~700 lines of code while also extending functionalities.

After this, gpu and cpu just create a gpu_device or cpu_device on the fly and apply it to the argument:

gpu(x) = gpu_device()(x)
cpu(x) = cpu_device()(x)

This also fixes the problem with transferring conv layers to amdgpu devices discussed in #2498 (comment)

TODO

• revisit documentation
• check that all functionality is retained

[codecov]

Codecov Report

Attention: Patch coverage is 45.45455% with 6 lines in your changes missing coverage. Please review.

Project coverage is 33.46%. Comparing base (31dccd1) to head (a216eb8).
Report is 1 commits behind head on master.

Files with missing lines	Patch %	Lines
ext/FluxAMDGPUExt/functor.jl	40.00%	3 Missing ⚠️
src/deprecations.jl	0.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2502      +/-   ##
==========================================
- Coverage   34.58%   33.46%   -1.13%     
==========================================
  Files          35       31       -4     
  Lines        1995     1829     -166     
==========================================
- Hits          690      612      -78     
+ Misses       1305     1217      -88     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[CarloLucibello]

The tests I temporarily marked as broken should be fixed by LuxDL/MLDataDevices.jl#88

[pxl-th]

Looks good to me (once broken tests are working again), especially since conv cpu-gpu transfer for AMD GPU is working again.

[CarloLucibello]

ops, should have waited LuxDL/MLDataDevices.jl#88 for unbreaking the tests. I'll wait for LuxDL/MLDataDevices.jl#88 before tagging.