Switch the POSIX HAL's mrb_hal_process_times() from times(2) scaled by
sysconf(_SC_CLK_TCK) to getrusage(RUSAGE_SELF)/getrusage(RUSAGE_CHILDREN),
converting each struct timeval directly into an mrb_process_clock_time.
This drops the clock-tick scale factor and reads CPU time to the
microsecond instead of a tick (usually 10ms).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VhjkXJBh21sZGNdh4gnufw

RUSAGE_SELF/RUSAGE_CHILDREN are XSI extensions, not guaranteed by base
POSIX.1, so a host without getrusage(2) now falls back to the previous
times(2)+sysconf(_SC_CLK_TCK) reading instead of failing to build.
Gated by a new MRB_PROCESS_HAVE_GETRUSAGE feature macro, following the
same #ifndef-overridable pattern as the other MRB_PROCESS_HAVE_* checks
in this file.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VhjkXJBh21sZGNdh4gnufw

assert_operator ..., :>=, ... passed even when Process.times always
returned 0.0, since 0.0 >= 0.0 is true. The busy loop between the two
readings burns far more CPU than any clock this gem falls back to can
fail to notice, so require a strict increase instead.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VhjkXJBh21sZGNdh4gnufw

times(2)'s return value is elapsed real time in clock_t ticks, not an
error code; POSIX overloads (clock_t)-1 for both. On a host with a
32-bit clock_t that value recurs by simple wraparound roughly every
497 days, well before any struct tms field would overflow, and a
raw equality check misreports that as a failure even though tms was
filled in correctly. Clear errno before the call and check it instead
of trusting the return value alone.

CRuby's Process::Tms is a Struct, and Struct#== checks the two
operands' classes for an exact match before comparing members --
unlike Process::Status#==, which reads a subclass by its values. The
previous is_a?(Tms) check let a Tms subclass with equal member values
compare equal, which Struct itself does not allow.

Process.times's cutime and cstime are POSIX's own count of waited-for
terminated children, credited whenever a wait(2)/waitpid(2) call reaps
one -- not only when this gem's own Process.wait or Process.waitpid
makes that call. mruby-io reaps children of its own (IO.popen(...).close,
say), and those show up in the totals too, so the doc comments in
process.c and process_hal.h no longer name only the two Ruby-level
methods as the source.

<sys/resource.h> was included unconditionally, so a build predefining
MRB_PROCESS_HAVE_GETRUSAGE=0 to force the times(2) fallback on a host
without getrusage(2) still failed to compile if that host also lacked
the header itself. Move the #include inside the same override guard so
overriding the macro skips the #include too.

Predefining MRB_PROCESS_HAVE_GETRUSAGE=0 was the only way to skip
<sys/resource.h> on a host that lacks it; without that override the
unconditional #include broke the build instead of falling back to
times(2). Probe the header with __has_include (GCC 5+/Clang/MSVC
19.15+) before including it, so hosts without it fall back on their
own; a compiler without __has_include keeps the old unconditional
#include, so those hosts still need the predefine.

The child previously ran "exit 0", too little CPU to distinguish real
getrusage/times(2) readings from a HAL that always answers 0, and the
assertions only checked non-decrease. Spawn a child that spins for a
measurable stretch instead, assert cutime/cstime are unchanged while it
sits unreaped, and that reaping it makes cutime + cstime increase
strictly.