Export/push behavior required repeating format/quantization across multiple calls, which could diverge (export() vs push()) and produce inconsistent outputs. This change introduces a single Turbo-level config for downstream export/push behavior with explicit defaults, format-safe quantization behavior, and v2.1 pre-release documentation/version updates.

• Shared Turbo-level export/push config

  • Added config support to turbo(...) and TurboModel.from_pretrained(...).
  • Introduced centralized defaults in TurboModel:
    • format: "safetensors"
    • push_format: "safetensors"
    • quantization: "Q4_K_M"
    • push_quantization: None
  • Added internal resolver logic to normalize and store this config once at initialization.
  • Supports nullable overrides for push_quantization (explicit None is preserved).

• export() now consumes internal config

  • format is optional; when omitted, uses shared config.
  • GGUF quantization defaults come from shared config (Q4_K_M) unless explicitly overridden.
  • Explicit format / quantization args still take precedence.

• push() now consumes internal config with format-safe quantization

  • format is optional; when omitted, uses push_format.
  • push_quantization is no longer always-on by default.
  • GGUF push still resolves quantization deterministically (falls back to shared GGUF quantization).
  • ONNX/MLX push no longer gets forced GGUF-style quantization when not configured.

• Docs + examples updated for new usage

  • Updated API/docs/examples to show configuring once at Turbo init, then calling export() / push() without repeating format/quantization in common GGUF flows.
  • Refreshed old usage snippets where the new shared-config API is preferred.

• Pre-release version update

  • Updated project/docs versioning and visible version strings to v2.1.0rc1 for pre-release preparation.

• Focused coverage for config propagation

  • Added tests validating:
    • deterministic default resolution,
    • export/push default propagation from shared config,
    • synchronization behavior when only export-side values are provided,
    • nullable push_quantization override behavior,
    • ONNX push does not force quantization by default,
    • portable temp staging paths via tmp_path.

from quantllm import turbo

model = turbo(
    "meta-llama/Llama-3.2-3B",
    config={
        "format": "gguf",
        "quantization": "Q4_K_M",
        "push_format": "gguf",
    },
)

model.export()               # uses gguf + Q4_K_M from shared config
model.push("user/my-model")  # uses gguf defaults from shared config