Inconsistent use of units in semantic conventions

[pirgeo]

What are you trying to achieve?
The semantic conventions seem to be inconsistent in their use of units.

Examples:
bytes, milliseconds and count are used by https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/metrics/semantic_conventions/rpc.md and https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/metrics/semantic_conventions/http-metrics.md.
By, s, 1 or {operations}, {faults}, etc. are used in https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/metrics/semantic_conventions/system-metrics.md

What did you expect to see?

According to https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/metrics/semantic_conventions/README.md#instrument-units, UCUM units should be used. The discussion in #705 seems to contradict that. To my understanding, the unit can be any free-form string, and the current spec supports that.

That means that all of the units in all of the semantic convention files linked above are valid. However, I think we should provide consistent units across the different semantic conventions, especially for units that are very common, like bytes and time units.