Avoid too large stack frames when using pbzero decoders

Before this CL TypedProtoDecoder ensured that the on-stack
storage is enough to cover the highest field id parsable by
protozero decoders. That lead to large stack frames of
1000 * 16 = 16KB per decoder, which in turn caused problems
in various internal usages (b/203389087, b/194510398).

This CL adds a little bit of extra complexity, allowing the
on-stack storage to be < than the highest known field id,
allowing to reduce the stack size by 10x, still reaping the
benefits of direct-on-stack decoding for most of the trace
protos.

Bug: 203389087

Test: manual:
  $ parallel -j10 -N0 time trace_processor_shell-before -q /dev/null mpts_trace ::: {1..10}
  $ parallel -j10 -N0 time trace_processor_shell-after -q /dev/null mpts_trace ::: {1..10}
  On a 266MB MPTS trace:
  Average: before: 14.6191, after: 14.6866
  Median:  before: 14.6385,  after: 14.5365

Change-Id: I8a51e11f1a1deb31df4bb02412d7fa0efb7277af