using tag_data_axes to transpose elements in dg kernel (instead of stupidly changing the kernel instructions themsleves)

allowing remove_work tags for dg generator (though current dg kernel construction function doesn't do anything with them)

allowing default tag values *different* from usual tag values to be used when user does not pass any tag vals

refactored generator (arg, value, type) lists as KernelArg class; now allowing a arg/tag to accept ANY_VALUE and use a default when no arg value is passed

added measurement kernel for the storing pattern of the dg_diff vec fetch pattern measurement kernel

moved code around for dg training kernels so that kernels used for training same variant are grouped together (one more)

moved code around for dg training kernels so that kernels used for training same variant are grouped together

updated access pattern version of lg overlap kernel to use lmem shuffle instead of local ops to avoid loop math optimizations

fixed weirdly fluctuating exec times in lg_overlap demo (possibly due to loop math optimizations) by changing local ops from math to shuffle