Rewrite the parsing engine for speed (3.7x rows, 10.6x with the new iter_terms)

[niconoe]

Rewrites the DwC-A parsing engine for speed, behind a characterization test suite built first.

Motivation: on a real import pipeline, parsing was about 70% of total time.

Results

Measured back to back on one machine, 400k rows, reading 14 of 50 columns:

Path	Time	vs before
Before	7.61s	1.0x
Row API (for row in dwca)	2.04s	3.7x
New iter_terms()	0.72s	10.6x

Opening an archive is now effectively free: the line offset index is built on first
random access rather than eagerly for every data file.

benchmarks/ holds the harness and the recorded before/after. Absolute timings vary
by up to ~40% between sessions on the same code, so only same-run ratios are meaningful;
the README says so explicitly.

What was actually slow

Not the CSV looping, and not the seek-based random access. Iterating an archive
instantiated a fresh csv.reader for every line and rebuilt the term-to-value
mapping per row. Together that was 78% of the profile. A control experiment (streaming
the file into the old Row class) gave no speedup at all, which is what pointed at the
real cause.

How it works now

• Iteration is a single streaming pass. Random access keeps its own path.
• The descriptor's field mapping is precomputed once per data file (FieldPlan).
• The offset index is lazy, byte-accurate, and indexes CSV records rather than
  physical lines, so a quoted field containing the line terminator no longer
  desynchronises it.
• DwCAReader.__iter__ returns a fresh generator, so iterations are independent.
• New opt-in iter_terms() yields a tuple per row for chosen terms, skipping the Row
  object and its dict. "id" and "coreid" request the key columns, using the same
  names headers() and pd_read() already use.

Behavior changes reviewers should know about

All are documented in CHANGES.txt. The engine ones were each pinned by a
characterization test that was deliberately inverted when the fix landed, so every one
is a visible diff rather than a silent change:

• ignoreHeaderLines above 1 skipped only one line when iterating a CSVDataFile,
  leaking header rows into coreid_index and orphaned_extension_rows.
• An undecodable byte desynchronised the offset index, silently truncating later rows.
• A field whose content started or ended with the fieldsEnclosedBy character had that
  character stripped.
• Nesting two loops over one reader ran the inner one once; calling get_corerow_by_id()
  inside a loop never terminated.
• A quoted field containing the line terminator was read as truncated by both access
  paths, with following rows misaligned.
• DwCAReader.next() now keeps an iterator independent of any for loop, and starts a
  new pass once exhausted. Iterate over the reader instead.

Also fixed, independent of the engine: headers dropped the column at index 0 for
metafile-less archives; hash() raised TypeError on every row; rows from two readers
over the same archive never compared equal; comparing a CoreRow to an ExtensionRow
raised AttributeError.

Packaging

dwca.star_record imported typing_extensions, which setup.py never declared, so a
pip-installed package raised ImportError. It now uses typing.Literal, which removes
the dependency and raises the floor to Python 3.8. The CI matrix drops 3.7 in the same
commit, and python_requires plus the classifiers now match it.

Testing

108 tests to 221. The suite was built before the rewrite specifically so the engine
could be replaced with evidence rather than hope, covering axes the existing fixtures
never touched: encodings, line terminators, quoting, header counts, ragged and blank
rows, extension files, and iteration semantics.

Two defects were found by fuzzing rather than by tests, because no sample archive in the
repo contains a newline inside a quoted field:

• iteration and random access disagreed on such archives, so CoreRow.extensions could
  return silently truncated data. A 400-archive differential fuzz (iteration vs random
  access vs a csv.reader ground truth) now passes 400/400.
• the record-aware index scanner was validated against csv.reader over 2000 generated
  archives at five chunk sizes before being written into the code.

Not included

• The version is not bumped and the changelog heading still reads
  v0.17.0 (unreleased) - cutting the release is the maintainer's call.
• pd_read() still does not forward encoding or quotechar, so a non-UTF-8 data file
  works through the row API but fails through pandas. Pre-existing, documented in the
  pandas tutorial, and a behavior change worth deciding separately.
• Reading data files directly from the zip without extracting (worth roughly 12%) was
  scoped out; it needs lazy per-member extraction to keep absolute_temporary_path()
  working.