perf: improve TTFX by type-erasing data movement

[Beforerr]

What changed

1. Type-erased data movement (src/loading/variable.jl). readblock! is now a thin wrapper; everything below it takes Ptr{UInt8} + element size + Vector{Int} dims/strides:
   • _read_records! — record range → bytes, threaded decompression compiled once per file format instead of per shape.
   • _copy_block! — one strided N-d gather kernel serving both the row→column majority swap and sub-block extraction.
   • Big-endian byte swap keyed on word size, not T.
2. VVREntry records per-leaf whether it's a CVVR. The old code took the last leaf's type for the whole variable, which is wrong when a compressed variable stores a poorly-compressing block as a plain VVR.
3. Array(var), collect(var), and show read via readblock! directly — DiskArrays' generic indexing costs ~60 ms compile per shape. Base.summary(::CDFVariable) is now cheap (Base's was ~190 ms/type).
4. Workload (src/precompile.jl): seeds one variable per common shape (floats ≤3-D, TT2000/Epoch 1-D, strings ≤2-D) with the ops people actually run — show, attribs, Array, var[1], record slice, DateTime.. Rare shapes skipped.

[github-actions]

Benchmark Results (Julia v1)

Time benchmarks

	main	e8dc10c...	main / e8dc10c...
elx/full_load	28 ± 1.4 μs	28.2 ± 1.5 μs	0.99 ± 0.073
elx/sum_array	2.2 ± 1.2 μs	2.14 ± 1.2 μs	1.03 ± 0.81
elx/sum_lazy	2.11 ± 1.2 μs	2.07 ± 1.2 μs	1.02 ± 0.85
elx/sum_var_access	2.47 ± 1.2 μs	2.32 ± 1.3 μs	1.06 ± 0.78
elx/var_access	0.257 ± 0.003 μs	0.248 ± 0.008 μs	1.04 ± 0.036
mms/full_load	14.4 ± 1.3 μs	14.8 ± 1.4 μs	0.974 ± 0.13
mms/sum_array	0.0904 ± 0.0036 s	0.0879 ± 0.0018 s	1.03 ± 0.046
mms/sum_slice	3.35 ± 0.36 ms	3.35 ± 0.082 ms	1 ± 0.11
mms/sum_var_access	3.36 ± 0.32 ms	3.34 ± 0.082 ms	1.01 ± 0.1
mms/var_access	0.094 ± 0.001 μs	0.096 ± 0.001 μs	0.979 ± 0.015
time_to_load	0.153 ± 0.0022 s	0.156 ± 0.0033 s	0.983 ± 0.025

Memory benchmarks

	main	e8dc10c...	main / e8dc10c...
elx/full_load	0.217 k allocs: 24.8 kB	0.217 k allocs: 24.8 kB	1
elx/sum_array	12 allocs: 28 kB	12 allocs: 28 kB	1
elx/sum_lazy	12 allocs: 28 kB	12 allocs: 28 kB	1
elx/sum_var_access	14 allocs: 28.3 kB	14 allocs: 28.3 kB	1
elx/var_access	1 allocs: 0.234 kB	1 allocs: 0.234 kB	1
mms/full_load	0.065 k allocs: 6.45 kB	0.065 k allocs: 6.45 kB	1
mms/sum_array	1.04 k allocs: 31.7 MB	1.04 k allocs: 31.7 MB	1
mms/sum_slice	0.069 k allocs: 1.29 MB	0.069 k allocs: 1.29 MB	1
mms/sum_var_access	0.069 k allocs: 1.29 MB	0.069 k allocs: 1.29 MB	1
mms/var_access	1 allocs: 0.219 kB	1 allocs: 0.219 kB	1
time_to_load	0.149 k allocs: 11.2 kB	0.149 k allocs: 11.2 kB	1

[codecov]

Codecov Report

❌ Patch coverage is 95.36082% with 9 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/loading/variable.jl	97.18%	4 Missing ⚠️
src/variable.jl	82.35%	3 Missing ⚠️
src/parsing.jl	86.66%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!