Add benches for data provider constructors

[sffc]

The data provider constructors are a big part of our startup cost. We should measure them and improve them if possible.

I think it's fine for the tests to measure the testdata versions of the providers.

[zbraniecki]

I have initial tests from my harness against datetime component, simple datetime formatting.
Params:

• style Medium, Medium
• locale "en"
• provider: StaticDataProvider (--all-keys, --all-locales): 12.4mb
• date constructed from epoch
• format to Rust string

Breakdown:

• ICU4C
  • constructor: 12.881 µs
  • format: 1.4387 µs
  • total: 27.600 µs
• ICU4X
  • provider: 42.943 µs
  • constructor: 6.1462 µs
  • format: 346.61 ns
  • total: 50.025 µs

Note: ICU4C total is larger than constructor+format which can be explained by caching. If I extract dtf constructor out of bench loop it probably performs some operations lazily on the first format and caches them.
ICU4X is more predictable - the total is the sum of 3 parts and extracting each one gives predictable total.

Summary - StaticDataProvider doesn't seem to be zero-cost and we should evaluate why and document it.

[zbraniecki]

(if we extract provider constructor, then ICU4C is 27us, and ICU4X is 6.4us)