Ryu-like String to Float Parsing (#138)

A limited implementation for up to 17-digit numbers.

See #129.

* Move some d2s intrinsics to d2s_intrinsics.h

* Rudimentary string to double parsing

* Implement more pieces for parsing

- parse exponents
- parse dot
- handle negative exponents
- handle rounding
- add some basic tests for these things

* Fix constants

* Use __builtin_clzll

Thanks to expnkx for finding this.

See #133.

* Allow exponent to start with '+'

* Correctly handle mantissa overflow

* Increase the table size

Denormal numbers require a larger table because the decimal exponent can go down to -326 (we're using the table inversely to how it's originally used for shortest).

* Handle most overflow/underflow situations

* Increase table size some more

* More tests close to the underflow

These are almost exactly halfway between 0 and the smallest representable float.

* Implement error handling

Mention the restrictions on the input, and that this implementation is experimental.

* Add another caveat to the documentation

* Make it compile with -DRYU_ONLY_64_BIT_OPS

This mode was not defining mulShift. We just use the same implementation as if it's not set.

* Use _BitScanReverse64 on Windows

* Rename mulShift to mulShift{32,64}

Since I moved the mulShift implementations to d2s_intrinsics, they conflict with the ones defined in f2s.c. This is only an issue when -DRYU_OPTIMIZE_SIZE is set, but I think it's better to disambiguate anyway.

* Rename max to max32 to avoid name conflicts on MSVC

* MSVC: Always use _BitScanReverse64

The HAS_64_BIT_INTRINSICS macro is false if RYU_ONLY_64_BIT_OPS is set, so using that is incorrect. Instead, only check for _MSC_VER. I can't seem to find a macro to check for existence of __builtin_clzll.