Avoid writing length on each IO for async channel

[vstakhov]

This is a simple fine tune for a metered channel to update channel length merely when we need a readout.

There are two other small changes:

• Implement len method for the futures_channel feature flag
• Fix inconsistency between documentation and the actual value for the TOF queue size

[drahnr]

I think once the root cause for some stalls are resolved, we should get rid of the two underlying implementations, and consolidate to one. Not a fan of the complexity added by sprinkled cfg(feature=..).

[vstakhov]

I have a gut feeling that we have deadlocks just because we rely somehow on a wild behaviour of the futures_channel. In fact, all the time we do clone for a Sender we have a free sending slot, so if we do clone+send it is guaranteed to be successful even if a queue is full. I personally would also prefer to get rid of code duplication and to use async_channel only (like Substrate does so far, for example).

[drahnr]

I am all in favor of consolidating to async_channel or whatever channel crate gets the job done without much headache.