Avoid C# GC when deserializing large messages

[jtattermusch]

Progress towards fixing #13236.

• add new SegmentDeserializer field into marshaller (that takes ArraySegment instead of byte[]). the backward compatibility is maintained as the new deserializer fallbacks to older byte[] (with some overhead though) based serializer if not defined explicitly
• when deserializing messages, don't always allocate a new byte array, but instead reuse a thread-local byte array that is allocated for that purpose. That byte array is always-growing, but it's a weak reference, so the runtime can free the memory if it grows too big and memory is needed (the expectation is users shouldn't really use messages larger >X MB or so at that makes little practical sense - they could always use streaming).
• new ParseFrom(buffer, offset, count) API from Protobuf 3.5+ is used to allow parsing messages from byte array that is potentially larger than the message size. Users would need to regenerate their .proto generate code to benefit from this optimization.

[grpc-testing]

****************************************************************

libgrpc.so

     VM SIZE        FILE SIZE
 ++++++++++++++  ++++++++++++++

  [ = ]       0        0  [ = ]


****************************************************************

libgrpc++.so

     VM SIZE        FILE SIZE
 ++++++++++++++  ++++++++++++++

  [ = ]       0        0  [ = ]

[grpc-testing]

[trickle] No significant performance differences

[grpc-testing]

[microbenchmarks] No significant performance differences

[grpc-testing]

****************************************************************

libgrpc.so

     VM SIZE        FILE SIZE
 ++++++++++++++  ++++++++++++++

  [ = ]       0        0  [ = ]


****************************************************************

libgrpc++.so

     VM SIZE        FILE SIZE
 ++++++++++++++  ++++++++++++++

  [ = ]       0        0  [ = ]

[grpc-testing]

[trickle] No significant performance differences

[grpc-testing]

[microbenchmarks] No significant performance differences

[grpc-testing]

****************************************************************

libgrpc.so

     VM SIZE        FILE SIZE
 ++++++++++++++  ++++++++++++++

  [ = ]       0        0  [ = ]


****************************************************************

libgrpc++.so

     VM SIZE        FILE SIZE
 ++++++++++++++  ++++++++++++++

  [ = ]       0        0  [ = ]

[grpc-testing]

[trickle] No significant performance differences

[jtattermusch]

CC @azyobuzin

[grpc-testing]

[microbenchmarks] No significant performance differences

[azyobuzin]

buffer.Length <= minSize should be buffer.Length < minSize

[azyobuzin]

It looks great as to deserializing mesages and works as I expect (with #13440 (comment) ). Thank you.

[apolcyn]

@jtattermusch is this waiting on review?

[jtattermusch]

@apolcyn I temporarily interrupted work on this PR to take care of other things (cmake & test flakiness). I'll get back to it when I'm back from vacation.

[jtattermusch]

Superseded by #18865 which provides a better solution.