Update GroupSizeFallbackMode logic

[nikita-savelyevv]

Changes

• Use ERROR mode by default, in the error message IGNORE/ADJUST modes are suggested to be used instead.
• Introduce group size value information to the bit-width distribution table.
• Set min_adjusted_group_size=32 by default instead of 16, because group size 16 usually results in lower performance compared to int8_asym.

Example compression log when group size is adjusted:

# nncf.compress_weights(
#     model,
#     mode=nncf.CompressWeightsMode.INT4_ASYM,
#     ratio=0.9,
#     advanced_parameters=nncf.AdvancedCompressionParameters(
#         group_size_fallback_mode=nncf.GroupSizeFallbackMode.ADJUST,
#     )
# )

INFO:nncf:Some nodes can't be quantized with the specified group size of 128. Adjusted group size values will be used.
Mixed-Precision assignment ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━100% 384/384 • 0:00:09 • 0:00:00
INFO:nncf:Statistics of the bitwidth distribution:
┍━━━━━━━━━━━━━━━━━━━━━━━━━━━┯━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┯━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┑
│ Weight compression mode   │ % all parameters (layers)   │ % ratio-defining parameters (layers)   │
┝━━━━━━━━━━━━━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┥
│ int8_asym, per-channel    │ 23% (224 / 385)             │ 10% (223 / 384)                        │
├───────────────────────────┼─────────────────────────────┼────────────────────────────────────────┤
│ int4_asym, group size 64  │ 2% (30 / 385)               │ 2% (30 / 384)                          │
├───────────────────────────┼─────────────────────────────┼────────────────────────────────────────┤
│ int4_asym, group size 128 │ 75% (131 / 385)             │ 88% (131 / 384)                        │
┕━━━━━━━━━━━━━━━━━━━━━━━━━━━┷━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┷━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┙
Applying Weight Compression ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━100% • 0:00:27 • 0:00:00

Reason for changes

User feedback.

Related tickets

175084

Tests

Updated existing tests.

[ljaljushkin]

Thank you, @nikita-savelyevv.
This PR should improve UX and stats look more clear now!

One minor suggestion to add , and - in the table:
int8_asym, per-channel
int4_asym, group size 32

[nikita-savelyevv]

Thank you, @nikita-savelyevv. This PR should improve UX and stats look more clear now!

One minor suggestion to add , and - in the table: int8_asym, per-channel int4_asym, group size 32

Thanks for the suggestion! Updated.