Hint to the compiler that returned indices are in-bounds

[coolreader18]

Closes #195

With this change, this function in a downstream crate:

pub fn foo(x: &[u8], y: u8) -> &[u8] {
    match memchr::memchr(y, x) {
        Some(i) => &x[i..],
        None => &[],
    }
}

Generates this assembly:

Assembly output

foobar::foo:
	push r14
	push rbx
	push rax
	mov eax, edx
	mov rbx, rsi
	mov r14, rdi
	lea rdx, [rdi + rsi]
	mov rcx, qword ptr [rip + memchr::arch::x86_64::memchr::memchr_raw::FN@GOTPCREL]
	mov rcx, qword ptr [rcx]
	mov edi, eax
	mov rsi, r14
	call rcx
	mov rcx, rax
	test cl, 1
	je .LBB0_1
	mov rax, rdx
	sub r14, rdx
	add rbx, r14
	jmp .LBB0_2
.LBB0_1:
	mov eax, 1
	xor ebx, ebx
.LBB0_2:
	mov rdx, rbx
	add rsp, 8
	pop rbx
	pop r14
	ret

Whereas before, it would include a branch to core::slice::index::slice_index_fail.

This only works for the oneshot functions, not memchr_iter, since Iter doesn't hold a reference to the original slice. To get the same effect there, I think you'd have to add a original_len: usize field to Iter, which isn't that crazy but maybe seems silly to only use it for an optimization hint.