rangefeed: limit memory budget used by a single tenant per kv node

[aliher1911]

Is your feature request related to a problem? Please describe.
Current rangefeed memory budget model has one pool for system tenant system ranges and another for all user ranges. This model while protecting kv nodes from running out of memory doesn't isolate tenants from each other. One tenant could theoretically hoard all rangefeed memory in a slow feed thus forcing other tenants to restart feeds and make catch up scans.

Describe the solution you'd like
A better approach could be limiting amount of memory used by each individual tenant per node. That would add another level in budget hierarchy where each range has a budget which is pooled with a capped budget per tenant which is then further pooled with all rangefeeds on the node.

Jira issue: CRDB-14456

Epic CRDB-37617

[erikgrinaker]

This should also have isolation between user and system ranges for the tenant, and be integrated with the new dynamic system span configs in #74555.

[aliher1911]

Do we want it as a single feature with all the follow ups for rangefeed budgets? I think having per tenant is pretty straightforward to implement while dynamic spans need some thinking as we need to come up with a way to handle dynamic info.

[erikgrinaker]

We can use a separate issue if you prefer.

[aliher1911]

Added #78977 for separate budgets and dynamic configs.