Index the destination once in _copy_subblocks!

[lkdvos]

_copy_subblocks! (used when converting a block tensor to a dense TensorMap, e.g. MPSKit's _fuse_env when preparing effective Hamiltonians) looped over every fusion-tree block of the destination and, for each, over every nonzero block of the source, reading it as v[f₁, f₂]. Each of those reads goes through subblockstructure → degeneracystructure, i.e. TensorKit's global LRUCache behind a lock, so the copy did (#fusion trees) × (#nonzero blocks) cache lookups.

This PR

• iterates the source blocks with subblocks(v) and indexes a single subblocks(tdst) iterator by fusion-tree pair, so each tensor's structure is looked up once;
• computes the block ranges from per-leg cumulative dimensions (one small vector per leg and sector, computed once), instead of a blockedrange per destination subblock.

Benchmark: copying the left environment × MPO block tensor of an N₂/cc-pVDZ DMRG2 step (D = 300, U₁×SU₂×fℤ₂, 416 nonzero blocks, 931 fusion-tree blocks in the dense result), single-threaded BLAS, results identical:

	serial	allocated	8 concurrent copies
before	0.537 s	0.107 GiB	2.18 s
after	0.237 s	0.068 GiB	0.386 s

The concurrent case is where the lock contention showed up: with several tasks preparing environments, the old version spent most of its time waiting on the cache's SpinLock.

Also bumps the version to v0.3.20. Full test suite passes locally (Julia 1.11.2).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
src/tensors/abstractblocktensor/conversion.jl	79.54% <100.00%> (-0.46%)	⬇️

... and 3 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.