fix: add support for wildcard pattern in seed dataset path

[nabinchha]

Fix for: #2

Users should be able point to a local folder with wildcard pattern like so (for parquet, json, jsonl, csv). If we follow this pattern, duckdb is able to read from all files across these extensions.

parquet_reference = LocalSeedDatasetReference(dataset="../my/long/path/*.parquet")
json_reference = LocalSeedDatasetReference(dataset="../my/long/path/*.json")
jsonl_reference = LocalSeedDatasetReference(dataset="../my/long/path/*.jsonl")
csv_reference = LocalSeedDatasetReference(dataset="../my/long/path/*.csv")

#8 should merge first.

We need this for BigIron to support any sort of non trivial partitioned seed dataset. We currently have a workaround in BigIron to consolidate partitions into one file, but that does not scale at all.

Example Preview result pointing seed dataset to "csv/*.csv"

[11:36:46] [INFO] 0️⃣ Using the first matching file in 'csv/*.csv' to determine column names in seed dataset
[11:36:46] [INFO] 🕵️ Preview generation in progress
[11:36:46] [INFO] ✅ Validation passed
[11:36:46] [INFO] ⛓️ Sorting column configs into a Directed Acyclic Graph
[11:36:46] [INFO] 🩺 Running health checks for models...
[11:36:46] [INFO]   |-- 👀 Checking 'nvidia/nvidia-nemotron-nano-9b-v2'...
col_names: ['language', 'greetings', 'name']
[11:36:47] [INFO]   |-- ✅ Passed!
[11:36:47] [INFO] 🌱 Sampling 10 records from seed dataset
[11:36:47] [INFO]   |-- seed dataset size: 22 records
[11:36:47] [INFO]   |-- sampling strategy: shuffle
[11:36:47] [INFO]   |-- selection: partition 2 of 3
[11:36:47] [INFO]   |-- seed dataset size after selection: 7 records
[11:36:47] [INFO] 📝 Preparing llm-text column generation
[11:36:47] [INFO]   |-- column name: 'greetings_completion'
[11:36:47] [INFO]   |-- model config:
{
    "alias": "nano-v2",
    "model": "nvidia/nvidia-nemotron-nano-9b-v2",
    "inference_parameters": {
        "temperature": 0.5,
        "top_p": null,
        "max_tokens": 2048,
        "max_parallel_requests": 4,
        "timeout": null,
        "extra_body": null
    },
    "provider": null
}
[11:36:47] [INFO]   |-- default model provider: 'nvidia'
[11:36:47] [INFO] 🐙 Processing llm-text column 'greetings_completion' with 4 concurrent workers
[11:37:01] [INFO] 📊 Model usage summary:
{
    "nvidia/nvidia-nemotron-nano-9b-v2": {
        "token_usage": {
            "prompt_tokens": 397,
            "completion_tokens": 5102,
            "total_tokens": 5499
        },
        "request_usage": {
            "successful_requests": 10,
            "failed_requests": 0,
            "total_requests": 10
        },
        "tokens_per_second": 380,
        "requests_per_minute": 41
    }
}
[11:37:01] [INFO] 📐 Measuring dataset column statistics:
[11:37:01] [INFO]   |-- 📝 column: 'greetings_completion'
[11:37:01] [INFO]   |-- 🌱 column: 'language'
[11:37:01] [INFO]   |-- 🌱 column: 'greetings'
[11:37:01] [INFO]   |-- 🌱 column: 'name'
[11:37:01] [INFO] 🎉 Preview complete!

[eric-tramel]

big fan, here. Thanks @nabinchha !

[nabinchha]

Found another validation fix we need to make while testing a notebook:
690b00f cc @johnnygreco

As is, we will not support this when we need to upload to a datastore.