Add zero layout to optimize alpha/beta=zero.

[maleadt]

Extracted from #113
Fixes #110

[maleadt]

Benchmark results for commit 1b2b59c (comparing to ac26708):

ID	before	after	change
["wmma", "Float16*Float16=Float32 (N=16384, A=T, B=T)"]	344.778 ms ± 3.101 ms	325.048 ms ± 8.137 ms	7.2% ✅

[maleadt]

Interesting, this doesn't seem the reason for the speed-up. Although this seems good to have anyway.

Before:

;  @ GemmKernels/src/kernel.jl:173 within `matmul_pipelined`
; ┌ @ GemmKernels/src/layout.jl:105 within `load`
; │┌ @ none within `vloada`
; ││┌ @ none within `macro expansion` @ GemmKernels/src/layout.jl:28
; │││┌ @ int.jl:295 within `div`
      %34 = sdiv i64 %33, 4
; │││└
; │││┌ @ LLVM/src/interop/pointer.jl:85 within `unsafe_load`
; ││││┌ @ LLVM/src/interop/pointer.jl:9 within `pointerref`
; │││││┌ @ LLVM/src/interop/pointer.jl:9 within `macro expansion` @ LLVM/src/interop/base.jl:39
        %35 = getelementptr inbounds <4 x float>, <4 x float> addrspace(1)* %25, i64 %34
        %36 = load <4 x float>, <4 x float> addrspace(1)* %35, align 16
; └└└└└└
;  @ GemmKernels/src/kernel.jl:175 within `matmul_pipelined`
; ┌ @ GemmKernels/src/layout.jl:114 within `store!`
; │┌ @ none within `vstorea!`
; ││┌ @ none within `macro expansion` @ GemmKernels/src/layout.jl:44
; │││┌ @ int.jl:295 within `div`
      %37 = sdiv i64 %24, 4
; │││└
; │││┌ @ LLVM/src/interop/pointer.jl:88 within `unsafe_store!`
; ││││┌ @ LLVM/src/interop/pointer.jl:46 within `pointerset`
; │││││┌ @ LLVM/src/interop/pointer.jl:46 within `macro expansion` @ LLVM/src/interop/base.jl:39
        %38 = getelementptr inbounds <4 x float>, <4 x float> addrspace(3)* bitcast ([0 x i8] addrspace(3)* @shmem to <4 x float> addrspace(3)*), i64 %37
        store <4 x float> %36, <4 x float> addrspace(3)* %38, align 16
; │││└└└
; │││┌ @ int.jl:86 within `-`
      %39 = add nsw i64 %24, 1024
; └└└└

After

;  @ GemmKernels/src/kernel.jl:175 within `matmul_pipelined`
; ┌ @ GemmKernels/src/layout.jl:114 within `store!`
; │┌ @ none within `vstorea!`
; ││┌ @ none within `macro expansion` @ GemmKernels/src/layout.jl:44
; │││┌ @ int.jl:295 within `div`
      %23 = lshr exact i64 %22, 2
; │││└
; │││┌ @ LLVM/src/interop/pointer.jl:88 within `unsafe_store!`
; ││││┌ @ LLVM/src/interop/pointer.jl:46 within `pointerset`
; │││││┌ @ LLVM/src/interop/pointer.jl:46 within `macro expansion` @ LLVM/src/interop/base.jl:39
        %24 = getelementptr inbounds <4 x float>, <4 x float> addrspace(3)* bitcast ([0 x i8] addrspace(3)* @shmem to <4 x float> addrspace(3)*), i64 %23
        store <4 x float> zeroinitializer, <4 x float> addrspace(3)* %24, align 16
; │││└└└