consistent KeyErrors across backends

[TeunHuijben]

Raise a consistent KeyError for missing attribute keys on all backends

Asking for an attribute key that doesn't exist raised a different exception depending on which
backend you were using:

	missing node key	missing edge key
SQLGraph	AttributeError	AttributeError
IndexedRXGraph	KeyError	ColumnNotFoundError
RustWorkXGraph	KeyError	KeyError (but ColumnNotFoundError via filter())

AttributeError leaks out of SQLAlchemy's getattr, and polars' ColumnNotFoundError isn't a
KeyError subclass, so except KeyError didn't catch every case. Code that worked against the
in-memory backends broke as soon as you switched it to SQL.

funtracks is the reason this goes first: its tests assert KeyError for a missing attribute key
and currently fail only on SQL.

What changed

All read paths now raise KeyError, with a message that names the unknown key and lists the valid
ones.

• Added BaseGraph._validate_attr_keys(attr_keys, mode). It's the read-path counterpart of the
  existing _validate_attributes, which guards writes and raises ValueError. Reads raise
  KeyError instead because asking for a key that isn't there is a lookup miss, like
  dict["missing"], while passing an undeclared key to a write is a bad argument.
• Called it from the read chokepoints in both backends: 4 places in _sql_graph.py, 6 in
  _rustworkx_graph.py. SQL needs fewer because all its reads funnel through one column resolver;
  rustworkx has separate implementations for graph-level edge reads, filter-level edge reads, and
  the neighbour walk in _get_neighbors.
• filter() validates at construction rather than at collect time, so a typo points at the
  filter() call instead of somewhere further downstream.

Covered: node_attrs, edge_attrs, nodes[id][key], edges[id][key],
filter(NodeAttr/EdgeAttr(...)), successors, predecessors, and the filter-level
node_attrs / edge_attrs.

While writing the tests I found one path that wasn't in the original bug report:
filter().edge_attrs raised ColumnNotFoundError on RustWorkXGraph too, not just
IndexedRXGraph.

Not in scope

Other backend inconsistencies we found but are handling separately: update_node_attrs silently
discarding a write with an unknown key on SQL, add_edge creating dangling edges on SQL, and the
differing exception types for a missing node/edge id.

Tests

• New parametrized test covering all 10 read paths × 3 backends.
• Tests for the error message content, and for filter() failing eagerly.
• A guard that valid keys are still accepted, including id columns, attr_keys=None, and
  attr_keys passed as a bare string.
• Tightened an existing test that accepted (KeyError, AttributeError) or no error at all.

Full suite: 1257 passed, 13 skipped, 1 xfailed.

Performance

No measurable cost. The check is one set lookup per call against the already-cached schema dict,
not per row. A/B on node_attrs over 20k nodes: SQL 15.4 → 14.6 ms/call, rustworkx
1.61 → 1.70 ms/call, both within run-to-run noise.

Breaking change

Anyone catching AttributeError for a missing attribute key needs to catch KeyError now.