cmov: use asm! barrier in portable backend (where stable)

[tarcieri]

On targets where asm! is stable but we don't have a backend using architecture-specific predication instructions, uses asm! to insert an fence/barrier which should hopefully accomplish the intended goal as well or better than core::hint::black_box, using a similar approach to what was introduced in zeroize in #1252.

Suggested by @newpavlov in #1332.

[tarcieri]

@NicsTr can you verify this still accomplishes the intended codegen?

[NicsTr]

When testing cmovnz with thumbv6m-none-eabi, it introduces a branch:

.section .text.not_ct::test_ct_cmov,"ax",%progbits
	.globl	not_ct::test_ct_cmov
	.p2align	1
	.type	not_ct::test_ct_cmov,%function
	.code	16
	.thumb_func
not_ct::test_ct_cmov:
	.fnstart
	.cfi_sections .debug_frame
	.cfi_startproc
	.save	{r4, r5, r7, lr}
	push {r4, r5, r7, lr}
	.cfi_def_cfa_offset 16
	.cfi_offset lr, -4
	.cfi_offset r7, -8
	.cfi_offset r5, -12
	.cfi_offset r4, -16
	.setfp	r7, sp, #8
	add r7, sp, #8
	.cfi_def_cfa r7, 8
	.pad	#4
	sub sp, #4
	uxtb r3, r2
	subs r2, r3, #1
	mov r4, r3
	sbcs r4, r2
	ldrb r2, [r0]
	mov r5, sp
	strb r4, [r5]
	@APP
	@ r5
	@NO_APP
	cmp r3, #0
	bne .LBB0_2
	mov r1, r2
.LBB0_2:
	strb r1, [r0]
	add sp, #4
	pop {r4, r5, r7, pc}

Is this approach of black_box (using inline asm), supposed to be more robust than the core::hint::black_box?

[tarcieri]

That was the hypothesis but perhaps it isn't true