[Python] Support tree-model TsFiles in TsFileDataFrame (#816)

* [Python] Support tree-model TsFiles in TsFileDataFrame

Squash of the tsdf-tree branch: tree-model support plus dataset catalog memory and indexing improvements.

- support tree-model TsFiles in TsFileDataFrame
- shrink the dataframe catalog memory footprint
- drop table-model phantom cells
- inline iter_owned_series_refs
- rename _LogicalIndex to _DataFrameCatalog
- dedup repeated series specifiers in tsdf.loc

* Fix tree-model metadata segments

* [Python] Dedupe dataset model constants and harden tree-model reads

- metadata: define MODEL_TABLE/MODEL_TREE once and import them in reader
  and dataframe instead of redefining the same literals in both modules,
  removing the drift risk between duplicate definitions.
- metadata: MetadataCatalog.series_count now returns the count of
  physically-present series (len(series_stats_by_ref)) instead of the
  device x declared-field cross-product, so it matches len(tsdf) and the
  reader property for sparse schemas; TsFileSeriesReader.series_count
  delegates to it as the single source of truth.
- reader: document why _metadata_field_stats gates on the value
  statistic's non-null row_count (not the timeline) and why the
  timeline_statistic reads below cannot raise.
- reader: mark the tree-model full-scan reads as O(total_rows) per device
  (cwrapper per-device pushdown limitation) so the hot path is easy to
  locate when profiling.
- reader: fail fast with an assertion if the cwrapper stale col_ leak
  pattern changes, instead of silently slicing the wrong path columns and
  returning wrong-device data.
- tests: add a sparse-schema regression assertion locking series_count to
  the physically-present count.

* [Python] Support unioning multiple tree-model TsFiles in TsFileDataFrame

- Tree TsFiles synthesize their schema per file, so loading several into one
  TsFileDataFrame previously raised "Incompatible schema" whenever the files
  differed in measurement subset or device depth. Widen the dataframe-level
  synthetic table to the union instead: _merge_tree_table_entries unions the
  field columns (existing order first) and grows the tag columns to the
  deepest file's depth so shallower devices pad with nulls.
- Key each logical series by the global field index resolved by measurement
  name against the merged table, while each shard keeps its reader-local
  field index for the read path. This keeps per-(device, field) ownership
  exact and avoids mis-mapping measurements across files. For table model the
  global and local indices always coincide, so behavior is unchanged
  (incompatible table schemas are still rejected).
- tests: add multi-file tree coverage -- identical-structure merge (shards and
  time bounds), field union across files, and differing-depth union with tag
  padding.

* [Python] Filter non-numeric measurements in tree-model TsFileDataFrame

The dataset surface is numeric (values are read as float64), and table mode
already drops non-numeric (STRING/TEXT) fields from its schema. Tree mode did
not, so a string/text measurement was surfaced as a series that then crashed on
read (float() on a string). Apply the numeric-type filter once in
_metadata_field_stats so both models behave consistently -- this also makes the
existing tree-model "measurements that pass the numeric filter" comment accurate.

- reader: skip measurements whose data_type is not numeric in
  _metadata_field_stats (shared by table and tree paths); table mode is
  unaffected since it already filters at the schema level.
- tests: add a tree-model regression test asserting a STRING measurement is
  dropped (not listed, raises KeyError) while the numeric one reads back.