proto_crypt: clear the remaining stack buffers

[woahwhattheheck]

Disclosure, per this repository's AGENTS.md: I am an LLM (Claude), submitting on behalf of the account owner. I am available to discuss this change and to revise it in response to review feedback.

Fixes #436.

#426 wiped dk_2 and nonce_y in proto_crypt_mkkeys(). Two buffers holding
key material in the same file were not covered by it:

• proto_crypt_secret() reads the key file into buf[BUFSIZ] and never zeroes
  it, on either the success path or the read-error path at err2. That buffer
  is the pre-image of the shared secret — the whole key file, for a key file of
  BUFSIZ bytes or less. The function already wipes the SHA-256 context two
  lines above, and proto_crypt_secret_free() zeroes the derived secret; the
  buffer it was derived from was the gap. Both spiped and spipe call this
  once at startup and then run for the lifetime of the process.
• proto_crypt_dhmac() leaves dk_1, the derived diffie-hellman MAC keys, in
  the frame when it returns. Called once per connection.

Three insecure_memzero() calls, using the idiom from #426.

Deliberately not covered

• ctx in proto_crypt_enc() and proto_crypt_dec() — it is a copy of
  k->ctx_init, which stays in the key structure for the whole connection, so
  wiping the copy buys nothing; and both functions run once per 1024-byte
  packet, which is not somewhere to add work for no gain.
• nonce_CS in proto_crypt_dhmac() — the nonces are sent over the wire in
  the clear.

Happy to add either if you would rather have the file uniformly wiped.

Testing

No behaviour change: the writes happen after the last read of each buffer in
every path, and neither buffer is read again. insecure_memzero.h is already
included by this file.