fix: Memory leak of g_tx_hash_ctx on aborted signing flows

handle_first_sign_chunk() allocates the global keccak context
g_tx_hash_ctx via APP_MEM_CALLOC and only the success path
(finalize_parsing_helper) and a subset of post-init parsing errors
(handle_sign:180) release it. Other early-error returns inside
handle_first_sign_chunk (init_tx failure, invalid data offset,
unsupported tx type, cx_hash_no_throw failure) propagate the SW
directly to handle_sign and out to main.c without freeing the
allocation. Likewise, reset_app_context() never released the context,
so app_quit() and the implicit reset at the top of
handle_first_sign_chunk left previous-session allocations dangling. A
host repeatedly triggering aborted INS_SIGN first-chunk APDUs could
exhaust the device's app heap.

Add a NULL-checked APP_MEM_FREE_AND_NULL of g_tx_hash_ctx to
reset_app_context(). Errors that bubble out of handle_sign reach
main.c with reset=true, which now centrally releases the context. The
existing late-error free in handle_sign is kept as defense in depth;
the NULL check makes the double path safe.

(cherry picked from commit d2829917d2fc9d0e39d760fcb6ef578af673f5c2)