perf(querier): classify time-series columns once per query

[tushar-signoz]

Description

Stacked on #12555 — the first commit here is that PR; review only the perf(querier) commit.

The graph reader decided what each result column was by switching on its Go pointer type, for every cell of every row — and the switch only listed 32- and 64-bit numeric widths. Anything narrower fell through and was silently dropped, so a raw ClickHouse panel aggregating or grouping by a small-integer or boolean column (max(severity_number), GROUP BY has_error, kind) either rendered an empty chart or merged every group into one unlabelled line.

• Each column is now classified once per query — timestamp, numeric, bool, string, or document — and every row just acts on that. Numeric coverage comes from the same helper that already counted numeric columns, so the two can no longer disagree and any width works.
• Rows reuse scratch instead of allocating per row: the aggregation slots, the label slices, and the label objects for rows that land in an already-seen series.
• Values skip the reflection boxing on the way out of the driver for the common types.
• The __result_<n> index is bounded before it sizes anything. The alias is user-written in raw SQL, and an overflowing or huge index (__result_99999999999999999999) crashed the reader with a panic — on main too, where the result buckets are sized from the same number after the rows are read. Indices past 1000 are now read as plain numeric columns.

Additional Information

Behaviour before/after, raw SQL panels on a 1M-row stack:

query	before	after
max(severity_number) over time (UInt8)	200 but an empty chart	values returned
grouped by severity_number	1 unlabelled series, 25 groups merged	25 labelled series
count() AS __result_99999999999999999999	panic	value returned

Cost, measured per request on a GROUP BY trace_id query returning 200k series, from pprof allocation deltas on a live server: 25% fewer bytes allocated and 23% less CPU (35% fewer allocations inside the reader itself). Response time is unchanged — these queries are dominated by ClickHouse and by encoding the response.