fix: prevent enabling ktls with a buffered record header fragment

[lrstewart]

Description of changes:

Minor ktls fix. We shouldn't enable ktls with ANY unhandled record data, including a partial header.
conn->header_in could contain data but conn->in not contain data if we blocked while trying to read the full record.

(Turning on ktls switches s2n-tls from parsing encrypted records to decrypted application data, so any partial data won't be handled right after the switch)

Callouts

There are a couple other places where we make similar checks:

• s2n_renegotiate_wipe, where we also check for "pending data" and fail if it exists. We already check header_in there.
• s2n_connection_release_buffers, where we check our IO buffers. However, not checking header_in is correct there. We're just releasing buffers for memory savings, and header_in doesn't alloc memory. Even if a partial header exists in header_in, we'll be just fine continuing the partial read that created it.

Testing:

Updated existing test

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[toidiu]

Can we just use s2n_stuffer_is_consumed? It would avoid the the == 0 and read better. Some of the tests could also be simplified.

Suggested change

	RESULT_ENSURE(s2n_stuffer_data_available(&conn->header_in) == 0, S2N_ERR_RECORD_STUFFER_NEEDS_DRAINING);
	RESULT_ENSURE(s2n_stuffer_data_available(&conn->in) == 0, S2N_ERR_RECORD_STUFFER_NEEDS_DRAINING);
	RESULT_ENSURE(s2n_stuffer_is_consumed(&conn->header_in), S2N_ERR_RECORD_STUFFER_NEEDS_DRAINING);
	RESULT_ENSURE(s2n_stuffer_is_consumed(&conn->in), S2N_ERR_RECORD_STUFFER_NEEDS_DRAINING);

[lrstewart]

The two methods are equivalent. However, from looking at how both are used, s2n_stuffer_is_consumed seems to be used when the memory associated with the buffer is the concern, and s2n_stuffer_data_available seems to be used when whether we can read from the buffer is the concern. From that perspective, we're using the right method here. Dropping "==0" isn't much of a gain, and s2n_stuffer_is_consumed doesn't simplify the tests at all.