Found by reviewing the module against CPython's `Modules/_zstd/` and the
pinned libzstd-rs-sys sources; each item below was reproduced before the
fix and re-checked after.

- `train_dict`/`finalize_dict` sized their output buffer with an
  infallible `Vec::with_capacity`/`vec![]` from a caller-supplied
  `dict_size`, so `train_dict([b'x'], 2**62)` aborted the interpreter
  instead of raising. Reserve fallibly and raise `MemoryError`, as CPython
  does.
- A digested dictionary was always built at `ZSTD_CLEVEL_DEFAULT`. Since
  `ZSTD_CCtx_refCDict` takes its parameters from the CDict, that silently
  discarded the requested level: `level=19` with `as_digested_dict`
  compressed a test corpus to 39817 bytes instead of 31951. Build the
  CDict at the compressor's effective level, tracked through `level=` and
  the `options` dict like CPython's `self->compression_level`.
- A bare `ZstdDict` was loaded as digested in both directions; CPython
  compresses with an undigested dictionary by default and only
  decompresses with a digested one. Give `DictLoader` a per-direction
  default.
- Neither compressor nor decompressor reset its session after an error,
  leaving the object permanently broken — every later call failed with
  "Operation not authorized at current processing stage" or re-reported
  the original corruption on valid input. Reset as CPython's error paths
  do, including restoring `last_mode` to `FLUSH_FRAME`.
- `apply_options` range-checked every parameter, rejecting values libzstd
  and CPython accept: 0 ("use the default") for `window_log`, `strategy`,
  `window_log_max` and friends, any non-zero value for the boolean flags,
  and the clamped `nb_workers`/`job_size`/`overlap_log`. Only
  `compression_level` needs the check — libzstd clamps that one silently,
  which is why `test_compress_parameters` requires a `ValueError` — so
  everything else now goes to libzstd, whose error code already maps to
  the same message.
- `ZstdDict` accepted content shorter than 8 bytes; CPython rejects it up
  front regardless of `is_raw`.
- `compress`/`decompress` refused `data` as a keyword and
  `get_param_bounds` required `is_compress` as one, both contrary to
  CPython's clinic signatures.
- The decompressor allocated a full 128 KiB scratch buffer even when
  `max_length` capped output far below that.

`test_zstd` (119 tests) and the snippet test pass; clippy is clean.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>