docs(lessons): add lesson 106 and trim three entries to budget

Lesson 106 records #1267's finding that an algebraically equivalent
rewrite of a float computation is bit-exact only when it preserves the
original's rounded intermediates: the issue's own resolution plan
proposed `partition_point(o < v + T)` as bit-exact for every input, and
an independent reviewer's 200 000-vector fuzz put it at 72 912
mismatches against the frozen scan.

Lessons 59, 84 and 89 were over the 75-line budget (92, 94, 87); each
now carries one sub-example per distinct mechanism (65, 65, 71), with
the remaining instances folded in as clauses and no citation dropped.

Refs #1267

Author: dekobon