Crypto: no unaligned 64-bit accesses on 32-bit ARM

[flatstik]

CRYPTOPP_ALLOW_UNALIGNED_DATA_ACCESS is defined whenever __ARM_FEATURE_UNALIGNED is set, which includes 32-bit ARMv6/ARMv7. There the hardware handles unaligned single-word loads and stores, but not LDRD/STRD/LDM, which the compiler uses for 64-bit words. With the macro defined:

• Camellia's camellia_load64()/camellia_store64() dereference uint64 pointers instead of using memcpy,
• Whirlpool's update passes unaligned input to HashMultipleBlocks() as uint64 * instead of copying it first,
• misc.h reports an alignment of 1 for every type.

On an OpenWrt ARMv7 (cortex-a15, musl) build, veracrypt --text --test stopped with SIGBUS in camellia_set_key() (ldrd r0, [r0]) from EncryptionTest::TestXts(), because the static test key happened to be only 2-byte aligned (0x57b9de). Whether it crashes depends on where the linker places the test vectors, so other builds pass by luck; real keys and data can be unaligned too.

This defines the macro for AArch64 only among the ARM targets (unaligned 64-bit accesses are fine there); 32-bit ARM then uses the memcpy/aligned-copy paths. x86, x64 and powerpc are unchanged; MSVC does not define __ARM_FEATURE_UNALIGNED, so Windows ARM builds are unchanged too.

Tested: the ARMv7 build passes the self-tests under qemu-user with -cpu cortex-a15 and -cpu cortex-a7 (before: SIGBUS on cortex-a15).

[idrassi]

Thank you for the fix.