Fix ln() rounding up on a result just below a rounding tie

[spokodev]

Problem

ln() is documented as correctly rounded, but for an argument whose natural
log lands just below a rounding tie it rounds up instead of down:

Decimal.set({ precision: 20, rounding: Decimal.ROUND_HALF_UP });
new Decimal('4.25914212183600678912449997353532582823393815').ln().toString();
// → 1.4490677601509316344   (should be 1.4490677601509316343)

The true value, computed at high precision, is
1.44906776015093163434999999999999999999999999975859147… — the 21st
significant digit is 4 followed by a long run of 9s, i.e. strictly below
the …3435 halfway point, so every half-rounding mode must round down to
…343. mpmath and this library at precision: 60 agree.

Affects all five HALF rounding modes (the default). Incidence on random input is
tiny but the failures are deterministic and reachable.

Root cause

naturalLogarithm and naturalExponential extend precision in a loop guarded by
checkRoundingDigits(…, repeating). When the argument-reduction reconstruction
lands the working value on a false exact-tie form (…5000 / …0000), the
repeating branch of checkRoundingDigits only detects a trailing …9999
run — unlike the non-repeating branch, which also detects …5000/…0000 ties:

// non-repeating branch already has:  … || rd == 50000 || rd == 0
// and:                               … || (rd == k / 2 || rd == 0) && … == 0

So the near-tie was reported as resolved and finalise rounded the false tie up.

Fix

Add the same …5000/…0000 tie/zero detection to the repeating branch, mirroring
the non-repeating branch (only naturalLogarithm/naturalExponential reach it).
Applied identically to decimal.js and decimal.mjs.

Verification

• The full test suite passes (22633/22633), including the added ln cases.
• A near-tie sweep of ln and exp across all five HALF rounding modes
  (20000 cases) against an 80-digit oracle reports 0 mis-roundings; several
  of the same inputs were mis-rounded before the fix.

Test

Added five correctly-rounded ln cases to test/modules/ln.js (results ending
…d 4999…). They fail on master (round up) and pass with the fix.

[MikeMcl]

Hi. I have seen this and will be on it shortly. Thank you.