SizeProfileComputer optimizations

- Refactor as an iterator: This allows filling ContentAnalyzerModule::aggregated_samples_
more directly and avoids redundant work.
- Use uint32_t field_idx/type as map key instead of string.
- Improve hashing function by shifting result.
- Use unordered_map instead of FlatHashMap: showed ~10% speedup in
  local runs.

Results:
On Linux cloudtop, chrome_android_systrace.pftrace-chrome ingest_time (% represents total overhead compared to baseline):
Baseline: 5.083
analyze enabled: 12.54 (146%)
patchset 1: 6.334 (24%)
+ accumulate sizes (using vector): 6.454 (27%)
+ reuse SizeProfileComputer: 5.92 (16%) <<<
+ using base::Hash: 6.15 (21%)

Change-Id: Ia46029466d62b29b24fea7ff0e2883501ab16b8b