DCompute: Inline memcmp for device array equality

[badnikhil]

Problem

The comparison/equality hooks added in #5149 lower array ==/< to a memcmp call (DtoArrayEqCmp_memcmp).That works on the host, but device targets have no libc to provide memcmp, and there is no llvm.memcmp intrinsic to fall back on ,so the call has nothing to bind to on the device.

Explanation of the Fix

• Added emitInlineMemcmp(), which emits a small loop (cond/body/diff/inc/end blocks) that walks the two buffers byte-by-byte and returns an i32 following C memcmp convention: 0 when all sizeInBytes bytes are
  equal, otherwise the unsigned difference of the first mismatching byte.
• callMemcmp() now branches on gIR->dcomputetarget: device code gets the inline loop, host code keeps the existing memcmp runtime call. The byte-size computation (numElements * elementSize) is shared by both
  paths.
• Modified the PHI in DtoArrayEqCmp_memcmp: because callMemcmp can now emit extra basic blocks, the merge phi takes its incoming value from the block we actually end up in (memcmpResultBB = GetInsertBlock()) rather
  than assuming the result is still produced in memcmpBB.

[thewilsonator]

test?