deflate should use __builtin_ctzll (64bit) instead of __builtin_ctzl (32bit)

[mgambrell]

I think the use of the 32bit operation is probably a mistake. It seems to leave an appreciable amount of compression on the table, perhaps by finding shorter matches (4 bytes) in cases where there may have been 5-8. I did not assess the performance implications. I did check the upstream code and the intention seems to have been to SWAR a fundamentally 8-byte serial process, but this intention was not achieved with the 32bit operation.

If my analysis is wrong for some arcane reason, then I think it warrants explaining in a comment at the point of the call why the strange choice was made.

[fhanau]

Nit: Since index is an unsigned long and at most 64, can you replace the return value here with unsigned long and remove the cast to int in the return expression? I doubt it matters for codegen but it should look slightly cleaner.

[fhanau]

Thank you for filing this PR. Please note that __builtin_ctzl takes in an unsigned long, which is 64-bit on many platforms (e.g. arm64 and x86_64 Linux and macOS), so in that case this works as intended.
On platforms where unsigned long is 32 bits (e.g. Windows), I think you're right and compression can be affected. When longest_match() finds that an 8-byte match is no longer possible, it tries to get a partial match of 0 - 7 bytes, but with a 32-bit __builtin_ctzl it would find at most 4 bytes, resulting in a shorter match length.
I left some comments on the __builtin_ctzll implementation for MSVC, LGTM otherwise. @vkrasnov @kornelski can you approve this once ready?