[CK_BUILDER] Add constraints for block tiling

[kabrahamAMD]

Added constraints for (c++20 concepts) for relations between thread block size, m/n_per_xdl and warp size for convolutions built with ck builder. In this process, it was also discovered that the example in test_conv_description.cpp was erronous and m/n_xdl_per_wave should be increased from 4 to 8.

[shumway]

This looks great!
That change to test_conv_description.cpp may break on gfx950, if so you may have to experiment a bit more with valid block tiling in that test.

[shumway]

Is this universal for all architectures (it looks like a generic mathematical condition)? Maybe add a short explanation why the conditions logically obvious or if there is a technical reference.

[vpietila-amd]

Would it be possible to break the constraints into more fine granular pieces to get a better compile-time error? I think in the current implementation it would be difficult to figure out which condition is violated.

[vpietila-amd]

The limits for tiles size are very useful since incorrect combinations of tile sizes are probably the biggest sources of errors when adding new kernel instances. However, I was under the impression that we wanted to collect the constraints into a single location where both the builder and the CK Tile implementation could use them. Maintaining a separate set of rules/constraints for the builder is a bit tricky since we don't have a good way of ensuring that they remain in sync with the implementation.

[shumway]

Looks like there's a failure on gfx942 with this test: experimental/builder/test/conv/test_ckb_conv_fwd_1d_i8.cpp

Those i8 tests seem to be the most fragile.