Compression latency needs a timer before the first image decode

An upload-start event emitted after image preparation cannot measure preparation latency. Whole photo-attempt durations are also unsuitable as compression benchmarks because camera interaction, user confirmation, upload, and validation can be included. Inspect timer placement before attributing a step latency jump to an encoder. Add monotonic decode/resize/encode timings, worker/fallback status, and an attempt identifier to separate these costs.