fix(training_shapes): add fallback passes and diagnostic logging for shape selection

[benjibc]

Problem

Training job rft--pyroworks-dev-ftsy30v1-e19defec7 for model qwen3p5-397b-a17b crashed because auto_select_training_shape couldn't find any matching shapes, even though validated shapes exist and are visible via firectl:

accounts/fireworks/trainingShapes/qwen3p5-397b-a17b-32k-b200-lora
accounts/fireworks/trainingShapes/qwen3p5-397b-a17b-262k-b200

The function's Pass 1 (exact base_model + trainer_mode server-side filter) returned zero candidates silently, then Pass 2 failed because the model API returned HTTP 403 for accounts/fireworks/models/qwen3p5-397b-a17b.

There was no logging to help diagnose why the combined server-side filter returned nothing — the shapes exist, are validated and public, yet the API listing returned empty.

Root Cause Analysis

The most likely causes for Pass 1 returning empty despite shapes existing:

1. Server-side trainer_mode filter mismatch: The API may store trainer_mode as a numeric proto enum (1, 2, 3) while we filter with the string name ("POLICY_TRAINER"). The combined server filter silently returns nothing.
2. Wildcard parent accounts/-/trainingShapes/- not listing cross-account shapes: On the dev gateway, the wildcard parent may not enumerate shapes from the fireworks account.

Fix

Added progressive fallback passes with diagnostic logging:

Pass	Strategy	What it catches
1	Exact base_model + trainer_mode server-side filter	Original behavior (unchanged)
1b	base_model only server-side, trainer_mode checked client-side	API doesn't support string enum filtering
1c	Account-scoped parent derived from base_model name	Wildcard parent doesn't list cross-account shapes
2	model_type + parameter_count bucket fallback	Different base_model name, same architecture
2b	Relaxed param filter (no trainer_mode server-side)	Same as 1b but for the param-count path

Also:

• Graceful 403 handling: _fetch_model_context failures are caught and logged, skipping the param-count fallback instead of crashing
• Comprehensive logging: Every pass logs its filter expression and candidate count at INFO level; _list_and_filter logs skip reasons (mode mismatch vs context too short) when API returns results but all are filtered client-side
• _account_from_base_model helper: Extracts the account from accounts/{account}/models/{model} for the account-scoped retry

Testing

• Added 23 new tests in test_training_shapes.py covering all passes, helpers, and edge cases
• All 354 existing unit tests continue to pass (331 passed + 23 new = 354; 32 pre-existing skips, 7 pre-existing failures from missing math_verify/eval_protocol deps)