For performance reason. Segmented list consumes slightly less memory
but takes a lot of time especially when there are many slots in the
segment lists (e.g. class constants).

Fix string test with 16-bit integers

Fix MSVC warnings in fmt_fp.c

fix fmt_fp.c(329) : warning C4244: 'return' : conversion from 'ptrdiff_t' to 'int', possible loss of data

fix fmt_fp.c(224) : warning C4244: '=' : conversion from 'ptrdiff_t' to 'int', possible loss of data

fix fmt_fp.c(206) : warning C4244: 'initializing' : conversion from 'ptrdiff_t' to 'int', possible loss of data

fix fmt_fp.c(178) : warning C4244: 'return' : conversion from 'ptrdiff_t' to 'int', possible loss of data

fix fmt_fp.c(123) : warning C4244: 'return' : conversion from 'ptrdiff_t' to 'int', possible loss of data

Fixed register windows of OP_SENDs generated during codegen

To enable method cache, define `MRB_METHOD_CACHE` or
`MRB_METHOD_CACHE_SIZE`. The cache size must be power of 2.
The default cache size is 128.

The measurement:
I measured simple benchmarks found in benchmark/ directory. With
method cache enabled, we gained 6-8% performance improvement, with
97-99% cache hit rate.

This only effects VC.

To calculate correct register windows size. The fix was suggested
by Christopher Aue.

Otherwise dead object may be seen from `#each_object`.