Improve objstore client for parquet Reading.

[cyriltovena]

This PR does fews things to improve the objstore client:

• On filesystem avoid to Open/Close/Stat a file for each ReadRange
• Instrument all clients
• Introduce a benchmark to verify the performance boost.

There's still probably a lot that we can do to ( cache footer and metadata informations) but for now this seems like a good step forward.

❯ go test -benchmem -run=^$  -bench ^BenchmarkDBSelectProfile$ github.com/grafana/fire/pkg/firedb -v -count=5 -timeout=0s -memprofile=mem.out -cpuprofile=cpu.out > after.txt && benchstat before.txt after.txt
name                old time/op    new time/op    delta
DBSelectProfile-16     123ms ± 4%      62ms ± 2%  -49.64%  (p=0.029 n=4+4)

name                old alloc/op   new alloc/op   delta
DBSelectProfile-16     114MB ± 0%     100MB ± 0%  -11.82%  (p=0.008 n=5+5)

name                old allocs/op  new allocs/op  delta
DBSelectProfile-16      478k ± 0%      281k ± 0%  -41.17%  (p=0.008 n=5+5)

[simonswine]

LGTM, see a couple of comments, also in need of rebasing

[simonswine]

Not a big fan of this name. Maybe we PersistentReaderAt or ReaderAtCreator is better

[cyriltovena]

Thinking about it.