feat: ✨ oversample SRC sketches before exact SVD truncation

[Copilot]

🤖 AI text below 🤖

Sketching at exactly chi_out makes randomized range-finding error dominate SRC accuracy. This change oversamples first, then truncates to the requested bond dimension.

• Sketch sizing: Add oversample to apply, compress, and src. Default to max(ceil(1.5 * chi_out), chi_out + 10); explicit factors use ceil(oversample * chi_out).
• Exact truncation: Add a backend-aware left-to-right SVD pass. Apply cutoff to Schmidt values in this pass and return left-canonical output. Store retained columns compactly.
• Compatibility: oversample=1.0 preserves the original sweep, cutoff semantics, and right-canonical output. The two-site exact fallback remains unchanged.
• Documentation: Describe the default, canonical forms, cutoff semantics, and larger-sketch cost.

# Sketch with 128 columns, then truncate to bond dimension 64.
out = apply(mpo, mps, chi_out=64, oversample=2.0)

• Fixes Oversampled sketch followed by an exact SVD truncation #40

[github-actions]

Test Results

    5 files  ±  0      5 suites  ±0   4m 17s ⏱️ +40s
  374 tests +108    367 ✅ +108   7 💤 ±0  0 ❌ ±0 
1 790 runs  +522  1 775 ✅ +522  15 💤 ±0  0 ❌ ±0 

Results for commit 8c0663e. ± Comparison against base commit a996df3.

This pull request removes 11 and adds 119 tests. Note that renamed tests count towards both.

tests.test_stack ‑ test_apply_matches_src[A B-0.0]
tests.test_stack ‑ test_apply_matches_src[A B-0.5]
tests.test_stack ‑ test_apply_matches_src[A psi-0.0]
tests.test_stack ‑ test_apply_matches_src[A psi-0.5]
tests.test_stack ‑ test_compress_matches_src[A-0.0]
tests.test_stack ‑ test_compress_matches_src[A-0.5]
tests.test_stack ‑ test_compress_matches_src[psi-0.0]
tests.test_stack ‑ test_compress_matches_src[psi-0.5]
tests.test_stack ‑ test_src_two_site_stack_falls_back[A B C]
tests.test_stack ‑ test_src_two_site_stack_falls_back[A B psi]
…

tests.test_gpu_backend ‑ test_oversampled_truncation_matches_cpu[mpo-1.0]
tests.test_gpu_backend ‑ test_oversampled_truncation_matches_cpu[mpo-2.0]
tests.test_gpu_backend ‑ test_oversampled_truncation_matches_cpu[mpo-None]
tests.test_gpu_backend ‑ test_oversampled_truncation_matches_cpu[mps-1.0]
tests.test_gpu_backend ‑ test_oversampled_truncation_matches_cpu[mps-2.0]
tests.test_gpu_backend ‑ test_oversampled_truncation_matches_cpu[mps-None]
tests.test_stack ‑ test_apply_matches_src[A B-0.0-1.0]
tests.test_stack ‑ test_apply_matches_src[A B-0.0-1.5]
tests.test_stack ‑ test_apply_matches_src[A B-0.0-None]
tests.test_stack ‑ test_apply_matches_src[A B-0.5-1.0]
…

♻️ This comment has been updated with latest results.