Add benchmark with a massive model variable

[michaelosthege]

The benchmark should recreate the shapes and dtypes from the following anonymized example:

<xarray.Dataset>
Dimensions:    (chain: 2, draw: 5000, a: 114574, b: 3)
Coordinates:
  * chain      (chain) int64 0 1
  * draw       (draw) int64 0 1 2 3 4 5 6 ... 4994 4995 4996 4997 4998 4999
  * a          (respondent) int64 0 1 2 3 4 5 ... 20819 20822 20823 20824 20825
  * b          (race) object 'first' 'second' 'third'
Data variables:
    v1         (chain, draw) float64 ...
    v2         (chain, draw, a) float64 ...
    v3         (chain, draw) float64 ...
    v4         (chain, draw) float64 ...
    v5         (chain, draw) float64 ...
    v6         (chain, draw) float64 ...
    v7         (chain, draw, b) float64 ...
    v8         (chain, draw) float64 ...
    v9         (chain, draw) float64 ...
Attributes:
    created_at:                2022-02-03T22:25:37.726525
    arviz_version:             0.11.4
    inference_library:         pymc3
    inference_library_version: 3.11.4
    sampling_time:             9668.074831008911
    tuning_steps:              1000

Note that the sampling time of this real world example is long enough to be an interesting use case for the ClickHouseBackend.