fix issues with olga heavy/light correction functions

[psathyrella]

Fixes #40.

1. In correct_olga_{heavy,light,paired}, rename the four
   qm.seq_gen_model (and seq_gen_model_{heavy,light}) assignments
   to match the qm.seqgen_model attribute the samplers actually read.
   See correct_olga_{heavy,light} doesn't reach the sequence generator #40 for a reproducer showing that the pre-fix writes land on a
   dead attribute so qm.seqgen_model.CPVJ never picks up the
   correction.

2. In correct_olga_light (and correct_olga_paired, light half),
   replace the open-ended [:, 1:] / [:, 4:] column slice with an
   explicit [:, [1, 2]] / [:, [4, 5]] so the unpack still works
   when Sonia.__init__ has added a trailing sequence_id column.

3. Persist the correction through save_model / load_model. When
   correct_olga_* has been called (tracked via a new
   qm._olga_corrected flag), _save_pgen_model now writes the
   corrected @v_choice (VDJ) or @v_choice + @j_choice (VJ) IGoR
   blocks into the saved model_marginals.txt instead of copying the
   OLGA defaults verbatim. Instances without the flag save
   byte-identically to before.

Test additions (all fail on the pre-fix source, pass after):

• test_correct_olga_light_no_shape_error /
  test_correct_olga_paired_no_shape_error — cover the shape fix.
• test_correct_olga_{heavy,light}_rebuilds_seqgen_model — cover the
  attribute-name fix by checking qm.seqgen_model and its CPVJ/CPV
  actually change.
• test_correct_olga_light_biases_sampled_v — end-to-end check that
  post-correction sampling follows the biased V/J.
• test_save_load_preserves_light_correction /
  test_save_load_preserves_paired_correction — round-trip via
  save_model + reload.
• test_save_no_correction_byte_identical — baseline guaranteeing
  models without a correction still save byte-identically to the
  shipped defaults.