Support proper numpy integration for ~100x performance boost

[VeaaC]

flatdata-py performance: vectorized access and scalar optimization

What

Adds NumPy-based vectorized field access to flatdata-py and optimizes the scalar (element-by-element) read path. Also fixes a pre-existing bug in read_value() for unaligned 64-bit fields.

Changes

Vectorized access (data_access.py, resources.py)

• read_field_vectorized(): reads a bit-packed field from all vector elements at once via NumPy, returning an ndarray. Zero-copy over the mmap'd buffer.
• Vector.__getattr__("field") returns a DataFrame column for the field.
• Vector.to_numpy() / to_data_frame() return all fields at once.
• _VectorSlice gets the same vectorized methods.
• Results are cached per vector instance via _as_numpy_2d().

Pre-computed field readers (data_access.py, structure.py)

• make_field_reader(offset, width, signed) builds a specialized closure with all constants (byte offset, bit shift, mask, sign handling) pre-computed. Six variants cover the cross-product of field types.
• Structure.__init_subclass__ builds a _READERS dict once per class.
• __getattr__, as_dict, as_list, as_tuple, as_nparray all use _READERS.
• read_value() is preserved as a thin wrapper around make_field_reader for one-off reads.

Bug fix (data_access.py)

• read_value() for 64-bit fields at non-byte-aligned offsets could return values wider than 64 bits (Python arbitrary-precision ints). The bit mask was only applied when num_bits < 64, missing the case where offset_extra_bits > 0. Fixed by masking when num_bits < 64 or offset_extra_bits > 0.

Other

• __slots__ = () added to generated Structure subclasses (generator template + 10 golden files). Reduces instance size from 72 to 48 bytes.
• Vector.__iter__ uses local variable caching to avoid repeated attribute lookups.
• Removed unnecessary list() on dict keys in Archive.__getattr__.
• Performance tips section added to flatdata-py/README.md.
• Version bump: flatdata-generator and flatdata-py both 0.4.10 → 0.4.11.
• CI workflow updated to install local generator before flatdata-py (py.yml).

Performance

Measured on a vector from a test archive (5.8M elements, 20 fields, 32 bytes each):

Access pattern	Before	After
Scalar iteration (1 field)	9.7s	5.8s
Vectorized column access (1 field)	n/a	0.07s