Drop ASCII-art dividers, colourise example output, ASCII-only source

[sbryngelson]

Summary

Two cleanups to the examples' presentation and to source-text hygiene. No library behaviour changes — all edits are display strings, comments, and docstrings.

1. Examples: drop ASCII-art dividers, colourise output

• Removed "=" * N / "-" * N printed dividers and # ----/# ==== comment rules across the examples.
• Added small helpers in examples/_common.py, following the conventional logging palette:
  • head() — bold green section headers (INFO), replacing the old =-banner blocks.
  • aside() — dim explanatory notes, so commentary recedes behind the numbers.
  • showwarning override — yellow warnings (WARNING), with each distinct warning shown once per run (collapses DispatchFloorWarning's per-call-site repeats).
• All colour is tty-guarded and NO_COLOR-aware: piped output / redirected files stay plain text. The library's warning filters are untouched (only the render step is overridden).

2. ASCII-only source

• Replaced 853 non-ASCII symbols across 70 files (aneforge/, tests/, bench/, + 3 native .mm/.h) with ASCII equivalents: em/en dash -> -, →->->, ×->x, ᵀ->^T, L⁻¹->L^-1, ≤/≥-><=/>=, ✓->OK, ·->*, etc.
• All occurrences were in comments, docstrings, and string labels — none in executable logic.

Verification

• ruff check: passes (CI lint job green).
• python -m compileall / py_compile: clean across all source; native dylib rebuilds from the edited .mm/.h.
• tests/op_smoketest.py: 22/22 ops correct on the ANE (M5 Pro).
• Full corpus (pytest tests/, forked): 524/527 passed.

Note on the 3 non-passing tests (pre-existing, NOT introduced here)

• test_cross_compile.py::test_relu_rejected_below_the_mil_floor — deterministic pre-existing bug: cross_compile_check for a sub-MIL-floor family (h11) raises ValueError in _targets.py instead of returning False. Reproduces identically on the unmodified code (the diff to _targets.py is comment-only).
• test_train_cifar.py::test_conv2d_pad_grad_matches_torch, ::test_group_norm_train_grad_matches_torch — pass in isolation; only segfault (signal 11) during the full forked mass-run (torch/libomp under hundreds of sequential forks). test_train_cifar.py is unmodified by this PR.

Tracking fixes for these separately.