Implement bignum library at level of machine code for certain targets

The current version of the CakeML compiler has a bignum library
that is generic for all targets and is implemented as part of
`data_to_word`. Version 1 of the compiler had a better bignum library,
but was specific to the x86-64 architecture.

This issue is about using the version 1 bignum library in the latest
version of the compiler when compiling to x86-64. Concretely, this
might be done by having a macro instruction in the compiler's `asm`
language. The macro is only available for x86-64 and expands to the
custom x86-64 code for bignum arithmetic.

At first, the improved bignum library implementation would be available
for x86-64 only, but similar custom code for other architectures would
also be possible.