As discussed in #711, the
current implementation of sending data is suboptimal when vectored
I/O is not enabled: data frame's head is likely to be sent in a
separate TCP segment, whose payload is of only 9 bytes.

This PR adds some specialized implementaton for non-vectored I/O
case. In short, it sets a larget chain threhold, and also makes
sure a data frame's head is sent along with the beginning part of
the real data payload.

All existing unit tests passed. Also I take a look at the e2e
/~https://github.com/hyperium/hyper/blob/0.14.x/benches/end_to_end.rs
but realize that all the benchmarks there are for the case of
vectored I/O if the OS supports vectored I/O. There isn't a specific
case for non-vectored I/O so I am not sure how to proceed with
benchmark for performance evaluations.

* Update indexmap to version 2

* Update webpki-roots dev-dep to 0.25

* Update tokio-rustls dev-dep to 0.24

* Update env_logger dev-dep to 0.10

* Remove combined minimal-versions + MSRV check for now

This PR changes the the assign-capacity queue to prioritize streams that
are send-ready.

This is necessary to prevent a lockout when streams aren't able to proceed while
waiting for connection capacity, but there is none.

Closes hyperium/hyper#3338

Co-authored-by: dswij <dharmasw@outlook.com>

This change causes GOAWAYs to be issued to misbehaving connections which for one reason or another cause us to emit lots of error resets.

Error resets are not generally expected from valid implementations anyways.

The threshold after which we issue GOAWAYs is tunable, and will default to 1024.