Redundant zero-extension when converting an int16 load to an int32#

[dvulakh]

Example program:

type bigstring = (char, Bigarray.int8_unsigned_elt, Bigarray.c_layout) Bigarray.Array1.t

external unsafe_get_16 : bigstring -> pos:int64# -> int =  "%caml_bigstring_get16u_indexed_by_int64#"

let unsafe_get_16_as_32 t ~pos = unsafe_get_16 t ~pos |> Stdlib_upstream_compatible.Int32_u.of_int

cmm (post #3336):

 (>>s
   (<<
     (let ba_data/658 (load_mut int (+a t/655 8))
       (load_mut unsigned int16 (+ ba_data/658 pos/656)))
     32)
   32))

asm (post #3336):

movq	8(%rax), %rax
movzwq	(%rax,%rbx), %rax
movslq	%eax, %rax
ret

The movslq is a noop after the movzwq, because there is no point in sign-extending the bottom 32 bits of %rax if the bottom 16 were populated with a zero-extending load.

It is possible to write this simplification at the cmm level, though doing so requires confidence that load_mut at a small size always becomes a zero-extending load.