Implement SerializationContext buffer writer

[JamesNK]

Fixes #609
Fixes #608

@mgravell

[JamesNK]

Before:

          Method |     Mean |     Error |    StdDev |      Op/s | Gen 0/1k Op | Gen 1/1k Op | Gen 2/1k Op | Allocated Memory/Op |
---------------- |---------:|----------:|----------:|----------:|------------:|------------:|------------:|--------------------:|
 HandleCallAsync | 2.794 us | 0.0533 us | 0.0654 us | 357,894.2 |      0.2098 |           - |           - |               880 B |

After:

                   Method |     Mean |     Error |    StdDev |      Op/s | Gen 0/1k Op | Gen 1/1k Op | Gen 2/1k Op | Allocated Memory/Op |
------------------------- |---------:|----------:|----------:|----------:|------------:|------------:|------------:|--------------------:|
 PipelinesHandleCallAsync | 2.472 us | 0.0296 us | 0.0262 us | 404,545.4 |      0.0687 |           - |           - |               296 B |

[mgravell]

Wow, it looks like you implemented option 3; gotta say: I'm impressed. Way to over-deliver!

[JunTaoLuo]

Some minor comments

[jtattermusch]

Overall looks goods but had a few questions.

[jtattermusch]

I think upon completion we should check that we've actually written the right amount of bytes to match the announced message size, otherwise we will get a protocol error at the peer (and that's something that's potentially hard to debug?)

[JamesNK]

I chose not to because this code will only be overridden by frameworks. They should test they are doing the right thing themselves using unit/functional tests.

Checking every serialization feels wasteful.

[JamesNK]

I'm going to commit this as it is. If you think this is important, or there are bug reports around this issue after release, then it can be added later by wrapping the buffer writer returned.

[ejona86]

This seems to have broken compatibility with grpc-java. I think it is because the gzip stream isn't "finish"ed. We can discuss on grpc/grpc#20884.