sha1dc: avoid misaligned uint32_t load when unaligned access is allowed - #7361

SHA1DCUpdate's whole-block loop casts the caller-supplied input pointer directly to uint32_t* and dereferences it whenever SHA1DC_ALLOW_UNALIGNED_ACCESS is defined, which happens by default on any x86/x86_64 build (see the SHA1DC_ON_INTEL_LIKE_PROCESSOR detection earlier in the file). That pointer's alignment is only known to the caller of git_hash_update/git_hash_buf, so whenever a hashed buffer is left at a non 4-byte-aligned address (for example, an object whose header length isn't a multiple of four, hashed together with its content in one call), this line reads a uint32_t through a misaligned pointer, which is undefined behavior in C and gets flagged by UBSan's alignment check.

The fix only takes the direct-cast fast path when the buffer is actually 4-byte aligned, falling back to the existing memcpy-into-the-context-buffer path otherwise, exactly like the code already does for the very first partial block.

Verification: since ordinary x86/arm64 hardware tolerates unaligned loads at the instruction level, this bug doesn't corrupt hash output on typical hardware/optimization levels, so a plain output comparison can't detect it. I verified it with a standalone harness built with -fsanitize=undefined,alignment (and -DSHA1DC_FORCE_UNALIGNED_ACCESS on this arm64 machine, to force the same code path that's active by default on x86_64):

• Before this change: UBSan reports "load of misaligned address ... for type 'const uint32_t', which requires 4 byte alignment" on every 64-byte block processed by SHA1DCUpdate when the source pointer is offset by one byte.
• After this change: the same input produces no UBSan diagnostic.
• In both cases, and in a build that forces the always-safe aligned path (SHA1DC_FORCE_ALIGNED_ACCESS), the resulting SHA-1 digest is identical, confirming the fix only removes the undefined behavior without changing any output.

Also added a regression test in tests/libgit2/object/raw/hash.c asserting that hashing the same bytes through a 4-byte-aligned buffer and through a deliberately misaligned buffer produces the same digest. The full test suite (libgit2_tests) passes locally.