Return HTC_S_OVERFLOW if exceeded.

This is a quick-fix for #2624, where we close the connection if a client
exceeds MAX_FRAME_SIZE. More proper handling with draining the rest of
the frame for the specific frames types that allow it (rfc7540 section
4.2) to come later.

Fixes: #2624

Fixes #2617

Instead of allocating struct h2_sess out of the same workspace that is
used as our rxbuffer, we rather keep it on the stack in h2_new_session.

If the size of the initial h/2 request payload is close to half the size
of the workspace, we would end up exhausting the workspace in HTC_RxInit
after allocating our h2sess struct.

Spotted by Federico G. Schwindt

Fixes: #2619

Ref #2645 but cannot be the cause because workspace is just not
overflowed

Doesn't change the logic however.

Hopefully this...

Fixes	#2605

... to save stack space

Fixes #2643

Ref #2643

Yes, I know, this is far from perfect, but it does not require
any changes outside the signal handler and will hopefully simplify
error report handling significantly.

Ref #2643

Spotted by:	fgs

purports to say on the tin, and by removing the quantum-test
it trips over.

better sense of it now.

Ref #2622

running the transaction.

almost certainly an error, so make it one.

Fixes #2613