Added an option to use zero-copy marshaller for the gRPC read operation

[veblush]

gRPC client benchmark added an option to use the zero-copy marshaller based on the recent protobuf optimization which has been available from protobuf-java 1.36.0 and the new gRPC feature (grpc/grpc-java#8102) which will be available from grpc-java 1.39.0. This is expected to reduce two memory copies involved in protobuf-message deserialization step.

[veblush]

This optimization turns out to be able to reduce the user cpu time by ~20% from the benchmark to download 1GiB file 30 times through Directpath.

# Without a zero-copy marshaller
$ time taskset -c 0 java -Xmx4096m -jar $JAR --client=grpc --bkt=gcs-grpc-team-veblush1 --obj=1GB.bin --calls=30 --method=read --size=1048576 --dp=true
real    0m59.133s
user    0m44.767s
sys     0m7.977s

# With a zero-copy marshaller
$ time taskset -c 0 java -Xmx4096m -jar $JAR --client=grpc --bkt=gcs-grpc-team-veblush1 --obj=1GB.bin --calls=30 --method=read --size=1048576 --dp=true --zeroCopy=true
real    0m57.722s
user    0m34.180s
sys     0m8.128s

Flamegraphs w/ and w/o the zero-copy also shows that it removes the cpu-cost of copying buffers. Note that right three pillars shrinked into 1 pillar with the zero-copy marshaller.

Without a zero-copy marshaller

With a zero-copy marshaller

Code to be optimized by this change (in the red box)

[veblush]

@voidzcy Could you take a quick look at the change to make sure if new features are used as intended?

[voidzcy]

LGTM