Fix TA-Lib integration and clean up indicator library selection

[LRAbbade]

Addressing a few issues with the TA-Lib path (select_stockstats_talib=1):

• It didn't use the user's tech_indicator_list — MACD/RSI/CCI/DX were always hardcoded
• Output columns used rsi, cci, dx rather than the stockstats-style rsi_30, cci_30, dx_30, so switching libraries would break downstream code
• No NaN handling before TA-Lib calls, which doesn't tolerate gaps well
• Common indicators like Bollinger Bands and SMAs weren't implemented

This is a follow-up to FinRL PR #1344 (already ported in #353), which fixed the auto_adjust / OHL adjustment in YahooDownloader. That addressed the data ingestion side — this one covers the indicator computation layer.

Further improvements:

• Replaced the int flag with an IndicatorLib enum — IndicatorLib.STOCKSTATS / IndicatorLib.TALIB instead of 0/1.
• Split add_technical_indicator into separate methods per backend instead of one large if/else block.
• stockstats path now operates on a copy to avoid StockDataFrame.retype() mutating the original DataFrame.
• Exposed drop_na_timesteps through the public API in data_processor.py.

[LRAbbade]

Ok I understand this would cause backwards compatiblity issues, but I do think the legibility is significantly improved, so why don't we meet in the middle? Did a small update so it can still be called with 0,1, but the enum would also work.

from meta.data_processor import DataProcessor
from meta.data_processors._base import DataSource, IndicatorLib
from finrl.config import INDICATORS

p = DataProcessor(
    data_source=DataSource.yahoofinance,
    start_date='2025-01-01',
    end_date='2026-01-01',
    time_interval="1d",
)
p.download_data(ticker_list=['NVDA', 'MSFT', 'META'])
p.clean_data()
p.add_technical_indicator(INDICATORS, select_stockstats_talib=0)
p.dataframe

Example of old code using stockstats lib:

Example of old code using TA-Lib lib:

Example of new code using stockstats lib:

Example of new code using TA-Lib lib:

Note the difference in length of the TA-Lib result is because of dropping nans, which is why I exposed drop_na_timesteps in DataProcessor.add_technical_indicator

[BruceYanghy]

Thanks for the fixes — this addresses several real issues in the TA‑Lib path.

Key wins:

• TA‑Lib now respects the user-supplied tech_indicator_list instead of hardcoding a few indicators.
• Output column names align with stockstats conventions (rsi_30/cci_30/dx_30), which avoids downstream breakage when switching libs.
• NaN handling and common indicators (BBANDS, SMA) are added.
• Cleaner separation between stockstats/talib paths, with legacy 0/1 selection still supported.

I’m in favor of merging once CI is green. The only behavior change to be mindful of is the default drop_na_timesteps and the TA‑Lib indicator whitelist (e.g., rsi vs rsi_30). If CI passes, this looks good to go.