Allow mapping object store paths without going through inventory

[ozkatz]

Currently there are 2 ways of "importing" data into lakeFS without actually copying it:

1. using lakectl fs stage or the equivilent stageObject API endpont.
2. Using lakefs import that utilizes the S3 inventory to read an entire bucket (and potentially only load a subset into lakeFS)

While (1) provides a reasonable solution for importing a single object, reading a directory or common prefix requires scripting that may or may not be trivial for the user. On the other hand (2) is great for loading a big (>1M objects) bucket into lakeFS - the ops overhead is substantial because:

1. it's a lakeFS command, not a lakectl one - so the machine running it needs access to the PostgreSQL database, to the underlying storage, to the lakeFS config values as well as a lakeFS binary
2. It requires a functioning inventory (and a good understanding of s3 inventory) configured and accessible to the caller
3. It requres S3 - lakeFS doesn't support this with other object stores.

We're missing a middle ground - the ability to ingest data directly from the object store into lakeFS, by using native object listing. This provides a relatively easy way to load a common prefix, a small table or a set of partitions (<1M objects) in a way that is more accessible to a data engineer.