fix: return na for empty array calculations

[luisleo526]

Summary

• return typed numeric na for empty array sum/avg/min/max/range, stdev/variance, median/mode, percentiles, and covariance before unsafe STL access or division
• preserve Pine argument evaluation: stateful scalar arguments are bound exactly once before the empty guard, while temporary receivers remain receiver-first
• make covariance safe for temporary peer arrays and preserve standardize(empty) as an empty array

TradingView documents empty numeric array calculations as na, except standardize, which returns an empty array: https://www.tradingview.com/pine-script-docs/language/arrays/

This is KI-48 stage 1 only. Mixed-NA filtering, negative/OOB access, and insert/fill/slice range errors remain separate stages.

Regression evidence

• focused semantic and compile review: 28/28, independent GO
• full engine-enabled pytest: exit 0, 1,651 collected with only 2 pre-existing skips
• corpus transpile/compile: 254/254
• gate selftest: 18/18
• full Pyodide parity: 277/277
• full 796-source baseline/candidate census: 779 OK + identical 17 errors, 0 status changes, exactly 15 generated-output changes
• 5 affected public Excellent strategies: candidate engine tapes byte-identical
• 2 affected public corpus strategies: candidate tapes byte-identical and still Excellent
• git diff --check clean

Expected public score impact: neutral. This fixes latent wrong-value/undefined-behavior paths without changing any currently exercised nonempty result.