Improve error messages for insufficient future_covariates in backtest() and fix residuals()

[AddyM]

Fixes #2846

Summary

This PR improves error handling when backtest() is called with insufficient future_covariates, replacing cryptic IndexError and TypeError messages with clear, actionable ValueError exceptions. It also fixes a bug in the residuals() method where covariate parameters were not being properly forwarded.

Changes Made

1. Enhanced Error Messages in backtest()

**Added validation to detect when future_covariates are provided but do not cover the required time range. New error messages now include: **

• Specific problem description with timestamps
• Exact number of missing time steps
• Clear resolution steps with code examples
• Alternative approaches (encoders, relative index)

Before:
model.backtest(series, future_covariates=short_cov, forecast_horizon=6)
IndexError: index 1 is out of bounds for axis 0 with size 1

After:
model.backtest(series, future_covariates=short_cov, forecast_horizon=6)
ValueError: TFTModel requires future_covariates to extend beyond the series end.

 **Problem: Series ends at: 2020-04-09, Future covariates end at: 2020-03-15
                     Required: 2020-04-15 (series end + 6 steps) 
                     Missing: 31 time steps**

  **Solution: Extend future_covariates to cover the full series + 6 steps.**

2. Added Validation Helper Method
Created _validate_future_covariates_for_forecast() to validate:

• Covariates start at or before the series start time
• Covariates extend at least forecast_horizon steps beyond the series end
• Validation only occurs when covariates are provided (model handles None case)
• Works for both single series and sequences

3. Fixed residuals() Bug
The residuals() method accepted past_covariates and future_covariates parameters but didn’t forward them to the internal backtest() call, causing potential validation failures or incorrect results. Now fixed.

4. Design Decisions
Validation added only to backtest() (not historical_forecasts()) to preserve flexibility for internal calls (e.g., ensembles). Only validates when covariates are provided; lets model handle None case. No API or behavioral breaking changes.

Testing:

• Added a comprehensive test suite: test_backtest_error_messages.py
  • Covers 5 main scenarios: Missing covariates, IndexError → ValueError conversion, TypeError → ValueError conversion, Regression tests without future_covariates, and edge cases for exact boundary conditions.
• All tests pass locally:
  • 5/5 new tests pass
  • All existing backtest tests pass
  • All residual tests pass
  • No regressions detected