Fix Newton inverse subtraction precision

[OskarEichler]

Summary:

• remove an accidental nested SIZET2NUM conversion in Newton reciprocal iteration
• pass the intended n / 2 precision directly to BigDecimal_sub2

Why:
The current expression converts n / 2 to a Ruby Integer, then feeds that VALUE back to SIZET2NUM as though it were a native size_t. On tagged-integer builds this inflates the requested precision substantially and performs avoidable work.

Verification:

• full current suite: 266 tests, 13,616 assertions, 0 failures/errors/omissions
• RBS suite: 146 tests, 1,805 assertions, 0 failures/errors
• five controlled 50,000-digit reciprocal runs, 50 iterations each: baseline average 0.25346s; fixed average 0.20403s (about 19.5% faster)
• numerical accuracy invariant passed in every run

Compatibility:
The correction restores the precision value already expressed by the surrounding n / 2 calculation; results remain within the existing accuracy invariant.