ext/mysqlnd: Guard against short server scramble in sha256_password auth

[iliaal]

mysqlnd_sha256_auth_get_auth_data() XORs SCRAMBLE_LENGTH (20) bytes of the server-supplied scramble into the password without checking the scramble is at least that long, unlike the native and caching_sha2 plugins which reject a short scramble with CR_MALFORMED_PACKET. A server reporting a scramble shorter than 20 bytes shrinks the heap buffer the scramble is copied into, so the XOR reads past it. Add the same length guard the sibling plugins use.

[morrisonlevi]

The code looks okay but... how would we test it to be sure?

[iliaal]

Added a regression test (a fake_server.inc scenario + .phpt): the fake server sends an AuthSwitchRequest to sha256_password with an 8-byte scramble, and the client now fails with CR_MALFORMED_PACKET instead of running the 20-byte XOR over the short heap copy. The over-read itself reproduces under ASAN (mysqlnd_xor_string reading past the mnd_emalloc'd scramble buffer).

[kamil-tekiela]

#21711

[kamil-tekiela]

@bukka I leave it up to you to decide on this. It's basically identical to my PR. I don't actually know if it fixes anything or just swaps one error message for another.

[iliaal]

It does fix a heap over-read; the old test just never reached it, since without a public key the XOR doesn't run. With a key and a longer password, valgrind on the unpatched build reports an invalid read in mysqlnd_xor_string() past the 8-byte scramble. #21711 came first and covers the same read, so I'm fine closing this one if the test moves there.