fix: initialize PreTypes sentinel to fix uninitialized read on stray #endif

[Exberg]

Summary

An extra #endif (no matching #if) made the preprocessor's termination
handlers (DoEndif, and the same peek pattern in DoElse, DoElseif,
DoEnddo, DoEndprocedure, ...) read AP.PreTypes[AP.NumPreTypes] with an
empty stack — i.e. PreTypes[0], a sentinel slot that was never
initialized. Malloc1 wraps malloc without zeroing (tools.c), so the
!= PRETYPEIF branch depended on heap garbage:

• Valgrind reported Conditional jump or move depends on uninitialised value(s) in DoEndif (pre.c), exactly as in this issue's report.
• Depending on the garbage value, the stray #endif could also be silently ignored (the non-reproducible case mentioned in the issue).

Fix

Initialize the sentinel at allocation time in startup.c:

AP.PreTypes[0] = PRETYPENONE;

This mirrors the existing sentinel initialization of the sibling stacks
(PreIfStack[0] = EXECUTINGIF in setfile.c, PreSwitchModes[0] = EXECUTINGPRESWITCH in pre.c). AddToPreTypes copies 0..MaxPreTypes when
it grows the array, so the sentinel is preserved across reallocations. No
behavioral change for correctly nested input: every handler pushes before the
slot can be read, and every matching pop restores NumPreTypes == 0.

With the fix, the exact reproducer from the issue now fails deterministically
and identically on every run:

test242.frm Line 5 ==> #endif without corresponding #if
Program terminating at test242.frm Line 1 -->

Regression test

Added Issue242 to check/fixes.frm: a stray #endif must always be
diagnosed (assert preprocess_error?("#endif without corresponding #if")).

Test plan

Built locally on macOS arm64 (./configure --disable-float && make -C sources form; MPFR not available on this machine — see note below).

• Issue reproducer: diagnostic now deterministic across 5 repeated runs
• ./check.rb ./sources/form Issue242 — 1 test, 1 assertion, passes
• Full default suite: 543 tests, 54 failures — every failure is a float-feature test (#StartFloat, mzv_, Evaluate, chop, torat, ...), reproduced identically (same 54 tests, diffed) on a pristine master worktree built with the same --disable-float configuration. They are a pre-existing consequence of building without GMP+MPFR on this machine, unrelated to this change. GitHub CI builds with the full feature set.

Note for reviewers: I could not run the Valgrind matrix locally (Linux-only
runner feature); the fix removes the uninitialized read by construction and
the deterministic test covers the observable behavior.

Closes #242

[coveralls]

coverage: 65.037% (-0.01%) from 65.048% — Exberg:fix/issue242-endif-uninit-sentinel into form-dev:master