[AllocsToSLM] Add thread-specific offsets

[dchigarev]

The proper way to use SLM via memref.alloc() appears to be allocating the SLM chunk for all threads in the workgroup at once, then slicing it for each thread. This PR implements this logic.

Example:

// Before the pass
func.func @entry() {
  gpu.launch blocks(%bx, %by, %bz) in (%sz_bx = %c8, %sz_by = %c8, %sz_bz = %c1)
             threads(%tx, %ty, %tz) in (%sz_tx = %c2, %sz_ty = %c4, %sz_tz = %c1) {
    %slm = memref.alloc() : memref<16x32xf16>
    gpu.terminator
  }
  return
}

// After the pass
func.func @entry() {
  gpu.launch blocks(%bx, %by, %bz) in (%sz_bx = %c8, %sz_by = %c8, %sz_bz = %c1)
             threads(%tx, %ty, %tz) in (%sz_tx = %c2, %sz_ty = %c4, %sz_tz = %c1) {
    // Scale allocation size by the number of threads in the work-group (16 * 2 = 32; 32 * 4 = 128)
    // This 'alloc' is only called once per work-group
    %slm_root = memref.alloc() : memref<32x128xf16, 3>
    // Compute the subview for each thread
    %slm = memref.subview %slm_root[%tx * 16, %ty * 32] [16, 32] : memref<32x128xf16, 3> to memref<16x32xf16, 3>
    gpu.terminator
  }
  return
}