Fix error reporting and resource cleanup across sources and constructors

[filtede98]

This patch fixes several error reporting and resource cleanup defects across source constructors and crypto wrappers:

1. zip_source_file_common.c (zip_source_file_common_new):

   • When a file does not exist (!sb.exists) and cannot be opened for writing (e.g. read-only, non-zero offset, or length specified), zip_error_set(&ctx->stat_error, ZIP_ER_READ, ENOENT) was called on the internal ctx->stat_error, and ctx was immediately freed. The caller-provided error pointer remained unpopulated, leaving callers of zip_source_file_create with NULL and an uninitialized error reason. This now correctly sets zip_error_set(error, ZIP_ER_READ, ENOENT).
   • On all error unwinding paths (!ops->stat, !sb.exists, length overflow, and zip_source_function_create failure), zip_error_fini is now invoked on &ctx->stat_error and &ctx->error before freeing ctx.
   • In read_file under ZIP_SOURCE_FREE, zip_error_fini is now called on &ctx->stat_error and &ctx->error before freeing ctx.

2. zip_source_free.c (zip_source_free):

   • When destroying a source, zip_error_fini(&src->error) is now called before freeing src. src->error is initialized in _zip_source_new(), and any dynamically allocated error string from operations on the source was previously leaked upon freeing.

3. zip_source_buffer.c (buffer_new, _zip_source_buffer_new, read_data):

   • In buffer_new, if allocating buffer fails, zip_error_set(error, ZIP_ER_MEMORY, 0) is now called, consistent with the other allocation checks in the function.
   • On zip_source_function_create failure in _zip_source_buffer_new and under ZIP_SOURCE_FREE in read_data, zip_error_fini(&ctx->error) is now called before freeing ctx.

4. zip_string.c (_zip_string_new):

   • If allocating s->raw fails, zip_error_set(error, ZIP_ER_MEMORY, 0) is now called before freeing s and returning NULL (matching the s = malloc(...) check).

5. zip_winzip_aes.c (_zip_winzip_aes_new):

   • If _zip_crypto_pbkdf2 fails, sensitive key material in buffer and ctx is now scrubbed with _zip_crypto_clear(), ctx is freed, and zip_error_set(error, ZIP_ER_INTERNAL, 0) is reported, preventing sensitive data leakage and unpopulated caller errors.

6. zip_source_pkware_decode.c (trad_pkware_free):

   • zip_error_fini(&ctx->error) is now called before freeing ctx, matching zip_source_pkware_encode.c.

7. zip_source_window.c (_zip_source_window_new, window_read):

   • zip_error_fini(&ctx->error) is now invoked on error unwinding and under ZIP_SOURCE_FREE.

8. zip_source_crc.c (zip_source_crc_create):

   • If zip_source_layered_create fails, crc_context_free(ctx) is now called instead of raw free(ctx), ensuring ctx->error is finalized.

[DeNdlerX]

I ran into the zip_source_free() leak independently while running the test suite under ASan on Linux, so I can confirm that part of this PR.

On current main, read_incons and decrypt-wrong-password-pkware-2 fail with a LeakSanitizer report. ziptool cat prints zip_error_strerror(zip_source_error(src)) and then calls zip_source_free(src), and since the source's error is never finalized, the string is never freed.

cmake -B build -DCMAKE_C_FLAGS="-fsanitize=address -g"
cmake --build build
ctest --test-dir build -R 'read_incons|decrypt-wrong-password-pkware-2'

Both pass with this PR. Valgrind shows the same leak on a normal build.

I also ran the full test suite with this PR under ASan. The only remaining failures are four tests using tryopen, which has a separate leak in the test program itself; I'll open a PR for that.

(clang 21, Ubuntu)