Use DF's planner directly for expression (filter) parsing

Currently, lance implements its own parser/planner logic in planner.rs in order to convert filters from SQL to datafusion expressions. This means we have to make extensions to lance's planner/parser to expose new capabilities as DF releases them (e.g. https://github.com/lancedb/lance/issues/1115 ). It would be easier to reuse DF's planner/parser directly. I've started an experiment to do this here: https://github.com/lancedb/lance/compare/main...westonpace:lance:experiment/use-df-planner?expand=1

There are a few prerequisite tasks before we can complete this.

• https://github.com/lancedb/lance/issues/849 - Lance's planner is designed to enable pyarrow.compute.Expression -> string -> DF expression. So it makes some choices (like a default timestamp resolution of microseconds) that are different than datafusion's planner. If we can instead go pyarrow.compute.Expression -> substrait -> DF then we don't need to worry about this.
• https://github.com/apache/arrow-datafusion/issues/7249 - Lance's planner is capable of handling timestamp literals with a specified precision (to choose between second/millisecond/microsecond/nanosecond resolution). Datafusion's planner can't yet handle this. So we need to add this capability to DF if we want to avoid a regression.
• we need to investigate / confirm DF's capability to handle UDFs and UDTs. Example from @eddyxu : we might want “Select * From … Where image_size(img) >= (256, 512)”
• we need to somehow support expressions from duckdb. One example of a problem here is that duckdb has a function is_valid and datafusion expects IS NOT NULL (not a function). Substrait could be a solution here if we write a converter (there are some seeds for this).
• Datafusion does not support nested nested references. For example, x.y is supported but not x.y.z