Centralize mojo_benchmark documentation.

mojo/devtools/common/docs/mojo_benchmark.md

 # mojo_benchmark
 
 `mojo_benchmark` allows you to run performance tests for any Mojo application
 that participates in the [tracing
 ecosystem](https://github.com/domokit/mojo/blob/master/mojo/services/tracing/interfaces/tracing.mojom)
 with no app changes required.
 
 The script reads a list of benchmarks to run from a file, runs each benchmark in
 controlled caching conditions with tracing enabled and performs specified
 measurements on the collected trace data.
 
 ## Defining benchmarks
 
 `mojo_benchmark` runs performance tests defined in a benchmark file. The
-benchmark file is a Python dictionary of the following format:
+benchmark file is a Python program setting a dictionary of the following format:
 
 ```python
 benchmarks = [
   {
     'name': '<name of the benchmark>',
     'app': '<url of the app to benchmark>',
     'shell-args': [],
     'duration': <duration in seconds>,
 
     # List of measurements to make.
     'measurements': [
-      '<measurement type>/<event category>/<event name>',
+      {
+        'name': <my_measurement>,
+        'spec': <spec>,
+      },
+      (...)
     ],
   },
 ]
 ```
 
+The benchmark file may reference the `target_os` global that will be any of
+('android', 'linux'), indicating the system on which the benchmarks are run.
+
+### Measurement specs
+
 The following types of measurements are available:
 
- - `time_until` - measures time until the first occurence of the specified event
- - `avg_duration` - measures the average duration of all instances of the 
-    specified event
+  - `time_until`
+  - `time_between`
+  - `avg_duration`
+  - `percentile_duration`
+
+`time_until` records the time until the first occurence of the targeted event.
+The underlying benchmark runner records the time origin just before issuing the
+connection call to the application being benchmarked. Results of `time_until`
+measurements are relative to this time. Spec format:
+
+```
+'time_until/<category>/<event>'
+```
+
+`time_between` records the time between the first occurence of the first
+targeted event and the first occurence of the second targeted event. Spec
+format:
+
+```
+'time_between/<category1>/<event1>/<category2>/<event2>'
+```
+
+`avg_duration` records the average duration of all occurences of the targeted
+event. Spec format:
+
+```
+'avg_duration/<category>/<event>'
+```
+
+`percentile_duration` records the value at the given percentile of durations of
+all occurences of the targeted event. Spec format:
+
+```
+'percentile_duration/<category>/<event>/<percentile>'
+```
+
+where `<percentile>` is a number between 0.0 and 0.1.
 
 ## Caching
 
 The script runs each benchmark twice. The first run (**cold start**) clears
 caches of the following apps on startup:
 
- - network_service.mojo
- - url_response_disk_cache.mojo
+ - `network_service.mojo`
+ - `url_response_disk_cache.mojo`
 
 The second run (**warm start**) runs immediately afterwards, without clearing
 any caches.
 
-## Time origin
-
-The underlying benchmark runner records the time origin just before issuing the
-connection call to the application being benchmarked. Results of `time_until`
-measurements are relative to this time.
-
 ## Example
 
 For an app that records a trace event named "initialized" in category "my_app"
 once its initialization is complete, we can benchmark the initialization time of
 the app (from the moment someone tries to connect to it to the app completing
 its initialization) using the following benchmark file:
 
 ```python
 benchmarks = [
   {
@@ skipping 27 matching lines @@
 
 If no `--server-url` is specified, the script assumes that a local instance of
 the dashboard is running at `http://localhost:8080`. The script assumes that the
 working directory from which it is called is a git repository and queries it to
 determine the sequential number identifying the revision (as the number of
 commits in the current branch in the repository).
 
 For more information refer to:
 
  - [Catapult project](https://github.com/catapult-project/catapult)