Update API structure

• Fast on large files: memory-mapped I/O, lazy block-level
  decompression with caching, and typed columnar decoding (no per-sample
  boxing). Channels decode in parallel.
• Typed signals: samples come back as []float64, []int64,
  []uint64, []string or [][]byte — never []interface{}.
• Streaming: read whole channels, sample windows, or iterate chunk by
  chunk for progressive loading.
• Robust: no panics on malformed files (fuzz-tested); errors carry the
  block type and file offset.
• Writes MDF files too: a streaming writer for data loggers (bounded
  memory, crash-safe unfinalized layout, Flush() durability points) and
  a batch column API for exports, with optional DZ compression.
• Decodes CAN bus logs: point it at a .dbc (or use the one embedded
  in the log) and get physical signals instead of raw frame bytes.
• Sample reductions: reads precomputed min/mean/max aggregates, so a
  viewer can draw a multi-million-sample signal from a few hundred
  records.
• Supports sorted and unsorted files, compressed data (DZ deflate +
  transposition, inflated in parallel), data lists (DL/HL), MDF 4.2
  column storage (LD/DV/DI), variable-length data (VLSD/SD),
  invalidation bits, non-byte-aligned integer channels, all three
  channel-array (CA) storage types, the channel hierarchy (CH) tree,
  unfinalized logger files (standard finalization steps applied in
  memory), and the full set of conversion rules (linear, rational,
  algebraic formulas, value/range tables, text tables).