libbinder: RPC cap transaction size at 100KB

Why?
- Android code uses -fno-exceptions and generally doesn't check for OOM
  conditions (unlike the Linux kernel itself!). Even if we check for
  allocation success, a successful allocation here may mean even a 1
  byte allocation on another thread or by the server will cause a
  failure.
- kernel binder can have by default 1MB of concurrent transactions
  at a time. A transaction of size 100KB is already exceedingly
  dangerous to the runtime, since in a big process, this could cause
  other processes to reach the limit.

In the future, we could increase this cap (lowering is potentially
difficult) or make it customizable.

Bug: 167966510
Test: binderRpcTest, binderRpcBenchmark, binder_rpc_fuzzer
Change-Id: Ia215f1a00412654ce08e6bced14d4da4a0a46987