Optimizing masked loads in software pipelined loops (or that are functions of ID)

[krzysz00]

This is me scribbling down a pair of optimization notes:

The first is that, if we peel off the last iteration of gemm/convolution/... reduction loops where we'll need to do masked loads (ex. because the K dimension isn't known to be evenly divisible by the reduction tile size), we won't need to do any masking of the loads except in that peeled iteration. This costs us code size, but might be a win in perf - someone should go try it.

The second note is that, if the mask condition is a relatively clean function of the wave/workgroup/... IDs, we should consider doing if (__unlikely(needs masking)) { slow path } else { fast path } fairly early during codegen so most parts of the kernel don't have to pay for inefficient loads at the cost of code size.