grpc: Enable SharedWriteBuffer option

From this slack thread, the KAS team observes a 30% savings of allocation thanks to [With]SharedWriteBuffer option. This option was introduced in upstream. Internally, by default, grpc allocates a buffer per "transport". In a busy server, those data buffering takes a significant memory footprint. This option allows all connections to share the same buffer pool. A connection returns the memory back to the pool when done. It reduces the number of allocations as well as helps Gitaly reduce the memory footprint with no/little downsides.

This is a good hanging fruit we could apply to Gitaly.