test(coverage): wire recd group into COV_RECD block for *_rec.c branch coverage

[gburd]

What

Adds a COV_RECD block (default on) to test/coverage/run_coverage.sh that
drives the recd recovery-record test group driver-per-test, raising
BRANCH coverage of the four *_rec.c recovery-handler files — the biggest
remaining cold branch surface in the tree.

The __xxx_recover handlers each have redo / undo / abort / do-nothing
branches. op_recover (testutils.tcl) crashes the env at every log-record
boundary and replays recovery forward (redo) and backward (undo/abort) —
exactly the cold branches. The subset previously ran only recd001/002/016
on btree, so the hash and queue handlers were at 0% and most
btree/db redo/undo branches were cold.

Results (recd-driven branch %, gcc --coverage, -O0)

file	before	after
db/db_rec.c	14.6% (467)	16.4% (523)
btree/bt_rec.c	12.3% (290)	21.6% (512)
hash/hash_rec.c	0% (0)	24.2% (501)
qam/qam_rec.c	0% (0)	45.5% (261)

Validated end-to-end via run_coverage.sh (COV_RECD=1, EXITCODE=0): all
26 recd runs PASS, each well inside its 300s per-test timeout (slowest is
recd004 at ~84s). Whole block ~10 min.

The curated set (coverage-per-minute)

Driver-per-test (each in its own tclsh — several recd tests use conflicting
globals, e.g. recd006 sets kvals as a scalar while recd010 uses it as an
array, so they cannot share one interpreter; each also spawns recdscript.tcl
subprocesses that need orphan cleanup):

• recd002 (splits) × btree/hash/queue/recno
• recd005 (file-id reuse) × btree/hash/queue/recno
• recd006 (nested txn) × btree/hash, recd013 (cursor adjust) × btree/hash
• recd004 big-key→internal, recd008 deep/many-child, recd009 recnum,
  recd010 off-page-dup splits, recd014 queueext (the only thing that
  reaches __qam_delext_recover), recd016 checksum, recd017 crypto,
  recd018 checkpoint/commit, recd019 txn-id wrap, recd020 dirs,
  recd022 aborted-prepared page-alloc, recd023 reverse split,
  recd024 streaming partial-put overflow, recd025 TXN_BULK.

recd001/002/016 removed from COV_TESTS (superseded by the driver-per-test
block, which is more robust than the shared-tclsh loop).

Deliberately excluded (too slow for the marginal branches, -O0)

• recd001 (~6.5 min/method): the rest of the set already covers its per-op
  redo/undo branches — dropping it entirely moved the totals by ~1 branch.
• recd003 (~7.3 min, dup): does not unlock the still-cold
  __bam_root/irep/rcuradj or __db_ovref handlers (they need scenarios no
  bounded recd test reaches), so its 7 min buy only a few adjacent branches.
• recd015 (~4.6 min): prepared-txn branches already hit by recd002/recd022.
• recd007 (~4.3 min): db_rec create/rename branches already covered by the
  recd_compact + upgrade drivers already in the subset.

Recovery correctness

No bug surfaced — all forward+backward replays across every access method
verified clean (that is the good outcome; a redo/undo bug would have failed
op_recover's post-recovery verify).

Follow-up (not in this PR)

Still-cold handlers that need scenarios no bounded recd test reaches:
__bam_root_recover, __bam_irep_recover, __bam_rcuradj_recover,
__db_ovref_recover, __db_cksum_recover. These would need a targeted C
recovery driver (like the existing recd_compact one) rather than a Tcl test.

[gburd]

Superseded by coverage/recd-branches-final (rebased onto master with #84/#88; COV_RECD block preserved, recd001/002/016 correctly removed from COV_TESTS in favor of the driver-per-test block). Reopening.