GeezerVison

Problem: When a select's condition is a known non-zero constant, the result
is always ifTrue. OptimizeInstructions::optimizeSelect already handled
this when ifFalse is pure by returning ifTrue directly, but it gave up
when ifFalse has side effects, because the obvious fold
(block (drop ifFalse) ifTrue) reorders ifFalse before ifTrue. That
left select instructions in optimized output that could have been folded.

Fix: Extend OptimizeInstructions::optimizeSelect in
src/passes/OptimizeInstructions.cpp to also handle the impure-ifFalse
case when
EffectAnalyzer::canReorder(getPassOptions(), *getModule(), ifFalse, ifTrue)
holds. The fold emits (block (drop ifFalse) ifTrue), which runs
ifFalse's effects, discards its result, then returns ifTrue. Value
semantics match the original select when the condition is non-zero, and
effect semantics are preserved. When canReorder fails, the select is
left unchanged; folding that case would require a temporary local to hold
ifFalse's value across ifTrue.

Verification:

./check.py succeeds. 453 gtest cases pass; 1027 lit cases pass; the
2 unsupported lit cases are host-related. All tool-level suites pass.

timeout 300 ./scripts/fuzz_opt.py --binaryen-bin=./bin ran 963 iterations
in 5 minutes without finding a bug. --optimize-instructions was exercised
directly and transitively via -O1/-O3/-Oz on inputs including
select-containing modules; --fuzz-exec-before, --fuzz-exec-second,
and CompareVMs reported no divergences.

The two affected lit tests (optimize-instructions-gc-tnh.wast,
optimize-instructions-mvp.wast) were regenerated with
scripts/update_lit_checks.py. The new CHECK lines expect
(block (drop <ifFalse>) <ifTrue>) for the folded cases and leave the
select unchanged when canReorder is false.

Follow-up: A more general fold would handle the case where canReorder
fails, by introducing a temporary local to hold ifFalse's value across
ifTrue. That would remove every constant-condition select from optimized
output, at the cost of one local per such site. Left for a separate PR
since it changes local assignment.

Fixes #6930.