Latx fix config edge cases

[LaurenIsACoder]

Summary

Fix two LATX string-handling boundary bugs and add a regression test.

Problem

Two boundary bugs can be triggered by ordinary runtime inputs.

• target/i386/latx/latx-options.c:trim() skips leading whitespace and
  then unconditionally evaluates s + strlen(s) - 1. If the input is
  empty or contains only whitespace, this moves the pointer before the
  start of the buffer and may cause an out-of-bounds access.

  One practical trigger is a configuration entry with an empty value,
  for example:

  LATX_CLOSE_PARALLEL=
  

or

LATX_CLOSE_PARALLEL=

• target/i386/latx/latx-special-args.c:extract_filename() truncates a
  long file name with strncpy(), but does not explicitly append '\0'.
  If the source name is longer than the destination buffer, the later
  strchr(buffer, '.') call may continue scanning past the end of the
  buffer.

Root Cause

• trim() assumes that at least one character remains after leading
  whitespace is removed.

• extract_filename() assumes that strncpy() always leaves a
  NUL-terminated destination string, which is not true for truncated
  copies.

Fix

• Return early in trim() when the trimmed string is empty.
• Stop the trailing-whitespace scan at the beginning of the string.
• Handle NULL and zero-sized buffers in extract_filename().
• Force NUL termination after strncpy() before scanning for '.'.
• Add a regression test covering both cases.

Validation

Build

CC=cc CXX=c++ ./latxbuild/build64-dbg.sh

Build the regression test

cc
-DNEED_CPU_H
-DCONFIG_TARGET='"x86_64-linux-user-config-target.h"'
-DCONFIG_DEVICES='"x86_64-linux-user-config-devices.h"'
-fsanitize=address,undefined
-g -O1
-Ibuild64-dbg -I. -Iinclude -Itarget/i386/latx/include
$(pkg-config --cflags glib-2.0)
tests/latx-x86_64/latx-config-regression.c
-o build64-dbg/latx-config-regression
$(pkg-config --libs glib-2.0)

Run

LATX_AOT=0 ASAN_OPTIONS=detect_leaks=0
./build64-dbg/latx-config-regression

Result:

• 5/5 tests passed on the fixed build.

Before the fix

Running the same regression binary against the pre-fix tree reports an
AddressSanitizer stack-buffer-underflow in trim().

Note on LATX_AOT=0 and ASAN_OPTIONS

• LATX_AOT=0 keeps the validation on the runtime LATX path touched by
  this change, instead of mixing it with AOT behavior.

• ASAN_OPTIONS=detect_leaks=0 only disables LeakSanitizer's leak check.
  It does not disable AddressSanitizer's out-of-bounds detection.

  This was added because LeakSanitizer may fail in some environments
  with a runtime/tooling issue such as ptrace, which is unrelated to
  the bugs fixed here. The actual boundary errors are still detected by
  AddressSanitizer normally.