fix: read mclock64() in os.clock(), not clock64()

[mgrossmann]

Fixes #8.

Why

libc370 #49 landed as mvslovers/libc370@f6d1888, changing clock64() to seconds
resolution — the behaviour time64.h:45 had documented all along.

os_clock() divides its source by CLOCKS_PER_SEC (1000) to split it into whole seconds
plus a millisecond remainder, so it needs a millisecond source. Reading clock64()
there now produces a value 1000x too small. Modelling the exact arithmetic against a real
STCK-derived timestamp:

true seconds since epoch : 1777488851.576
mclock64() -> os.clock() : 1777488851.576   <- this PR
clock64()  -> os.clock() : 1777488.851      <- current main
ratio (unfixed / true)   : 0.001000

The build does not catch it. clock64_t and mclock64_t are both plain uint64_t, so
the resolution mismatch is silent — os.clock() keeps compiling, linking and returning a
plausible-looking number.

lua370 is exposed automatically: mbt's reusable workflows pin cc370 via
inputs.cc370_ref but clone libc370 at main HEAD with no ref, and the toolchain cache
key embeds that SHA.

What changed

One line, plus the comment above it:

-	c.u64 = clock64();
+	c.u64 = mclock64();

mclock64() stayed at milliseconds through #49. clock64() had exactly one production
use in this repo, so the rest of os_clock() is untouched.

The comment was rewritten to record why this is deliberately mclock64(), and its stale
crent370 / crent470 paths dropped in favour of naming the functions — the original
note about libc370's clock() stub being the underlying reason is kept.

Verification

• Object level: loslib.o's external reference goes TM64CLCK → TM64MCLK (decoded
  from the EBCDIC object); confirmed against a rebuild of the unmodified file.
• Clean rebuild against the post-#49 sysroot: make, make lib, make test all rc=0
  under -Wall -Werror.
• Sysroot confirmed to carry post-#49 libc370 (the rebuilt tm64clck.o is byte-identical
  to the copy inside the installed libc.a).

Not affected

os.date, os.time, os.difftime use time64() / gmtime64_r() / localtime64_r() /
mktime64() / difftime64(). #49 moved time64() in the same commit but deliberately
kept its value unchanged — it used to divide clock64()'s milliseconds by
CLOCKS_PER_SEC and now returns clock64()'s seconds directly. Two cancelling constants
were removed together, so no change is needed here.

Left open (see #8 for detail)

Neither is closed by this PR, and neither is in this diff:

• os_clock() still divides a millisecond source by CLOCKS_PER_SEC, correct only while
  that macro is 1000. #49 removed libc370's last reader of it, so this is now effectively
  the only place using it as a wall-clock scaler.
• os.clock() returns elapsed wall-clock time rather than CPU time as Lua specifies. The
  root cause is libc370's clock() stub returning -1; once that is implemented for real
  (TCB / TIMEUSED), os_clock() should move to the existing #else branch.

Ref: mvslovers/libc370#49