Remove secp256k1 32-bit specific code.

[roconnor-blockstream]

This all but eliminate any concerns about different representations of secp256k1 jet results between 32 and 64 bit platforms (which could lead to consensus failures).

This can be tested using, for example,

nix-build -A c --arg nixpkgs '(import <nixpkgs> {}).pkgsi686Linux'

[roconnor-blockstream]

Note, we can still build on 32-bit platforms. It just uses the struct-based implementation of wide multiplication with the "64-bit algorithms" rather than the specialized 32-bit algorithms.

[apoelstra]

With my copy of nixpkgs if I try to build with pkgsi686Linux I get errors of the form

secp256k1/int128_native.h:16:9: error: unknown type name 'uint128_t'
   16 | typedef uint128_t secp256k1_uint128;
      |         ^~~~~~~~~

[roconnor-blockstream]

@apoelstra I think you are going to have to add debugging pragmas to https://github.com/BlockstreamResearch/simplicity/blob/1ee7c32576404d56f3164260f66342eff69a4e3c/C/secp256k1/util.h#L89-L120 in order to figure out why that is getting run on pkgsi686Linux. It should only be used if defined(UINT128_MAX) || defined(__SIZEOF_INT128__).

[apoelstra]

Oh, I see. Probably I am forcing these to be on as part of my CI setup.

[apoelstra]

I assume it's because I have wideMultiply set to "int128". I will update my CI setup to not do that on 32-bit builds.

[roconnor-blockstream]

Yeah, if you are testing with USE_FORCE_WIDEMUL_INT128, then that is expected to fail on 32 bit platforms.

[apoelstra]

Also (new in this PR) SECP256K1_WIDEMUL_INT64 is disabled for all platforms, not just 32-bit ones.

I would suggest removing the int64 option from the nixfile or at least erroring out on it early. (Can be done in a followup.)