fix: refactor channel refresh

[daniel-sanche]

python-bigtable performs pre-emptive refresh of its grpc channel. This means it has a background task to automatically cycle out the grpc channel with a new pre-warmed one every ~45 mins, to avoid latency spikes when the channel eventually expires

The current implementation has issues:

• it calls transport.create_channel() with no arguments to create a new warmed channel, meaning channel customizations are lost in the new channel
• populating interceptors is a little complicated, and involves accessing internal variables

To fix this, this PR adds a new SwappableChannel class, that wraps a real grpc channel, but provides methods to swap out the channel internally. This allows the channel refrence used by the gapic layer to remain constant, and leads to less abstraction leakage (issue 2/3). The ReplaceableChannel can also store the arguments used to init the channel, and re-use the same arguments when creating new ones (issue 1)

I performed some benchmarks, and didn't see a performance regression from this change

Fixes #1094